Global South AI Safety Hackathon

Permissive Models, Unequal Risk: Auditing AI Identity-Document Forgery as a Systemic Infrastructure Risk

Buenos Aires, Argentina

Two 2021 breaches exposed the identity records — including ID photographs — of essentially all of Argentina (RENAPER, ~45M, attacker-claimed) and Brazil (the megavazamento, ~223M). Frontier text-to-image models supply the forgery half, recomposing leaked photos into credentials that defeat appearance-based KYC. Our …

Marco ético para IA aplicada a la preservación lingüística Guna de Panamá

Team CREA TEAM · Colombia

This document aims to develop an ethical, responsible, and participatory framework (CREA) for the application of Artificial Intelligence in the preservation and use of the Guna language of Panama. This need arises from technological inequalities between rural and urban areas, which limit the Guna community in its …

Ufakazi

Team Ufakazi · Cape Town

Ufakazi is an evaluation harness built to determine whether models are unfairly biased towards trusting testimonies presented in high-resource languages (specifically English). By controlling for confounders, we isolate the language bias of various current generation LLMs and show that they often unfairly discriminate …

Confidently Wrong: Measuring and Mitigating Calibration Risks in LLMs for African Languages

Team AIMS South Africa · South Africa

Large language models are increasingly used in African languages, but little is known about whether their confidence remains trustworthy. We evaluate four open-weight LLMs across African-language truthfulness and knowledge benchmarks and find a consistent safety risk: models become less accurate while remaining highly …

When Safeguards Stop at the Border, Auditing How OpenAI and Anthropic Allocate Privacy Protections Across Latin American Jurisdictions

Team Los patiperro · Santiago, Chile

Analysis of AI privacy policies for LatAm countries based on their personal data protection laws and their comparison with foreign standards using a tool based on RAG architecture and with a judge based on LLM.

Prompteus: Confidential Third-Party Safety Auditing with Garbled Circuits, MPC, and PIR

Team paxus · Bangaluru

Regulators and downstream deployers in the Global South increasingly have to vouch for AI models they can't inspect, using evaluation data that can't legally leave the country. it lets a third party verify a safety property of a closed model without the owner revealing internals, without data contributors revealing …

IndicViet-Safe: Cross-Lingual Safety Evaluation of Open-Source LLMs in Hindi, Hinglish and Vietnamese

Team Suraksha · Gurgaon

We built a multilingual safety benchmark of 210 prompt-language pairs across English, Hindi, Vietnamese, and Hinglish (code-switched) covering 12 harm categories — including India-specific (caste discrimination, communal incitement) and Vietnam-specific (political sensitivity, censorship circumvention) threats. …

STEER - Mech interp-based white box attack on LLMs

Team JvThunder · Singapore

STEER (Safety Targeted Embedding Exploit via Refinement) is a white-box jailbreak that turns a mechanistic-interpretability finding into an attack: since safety fine-tuning encodes refusal as a single linear direction in the residual stream, STEER reads that direction, uses gradient attribution to pinpoint the words …

Do Multilingual Vision-Language Models Abstain under Cross-Modal Conflict in Low-Resource Languages?

Team Team_AVA · Hyderabad, India

When a VLM faces conflicting visual evidence and textual claims, like a red car captioned as blue, it must arbitrate between the modalities. While models sometimes safely abstain in English, we show this safety behavior fails in low-resource languages. We released a multilingual counterfactual benchmark covering …

Thought Anchors for Social Bias: Which Reasoning Steps Matter in Extended Thinking LLMs on Latin American Scenarios

Bogota, Colombia

We investigate which reasoning steps in extended-thinking LLMs are associated with pro-stereotypical outputs on Latin American social bias scenarios. Social bias benchmarks, including BBQ \citep{parrish2022bbq}, SESGO \citep{robles2024sesgo}, and EsBBQ/CaBBQ \citep{ruizfernandez2025esbbq}, measure final-answer …

The Shapes of Bias in Spanish-Prompted LLMs and the Debiasing Prompt Scaffolds

San Francisco

Largelanguagemodelsencodeandamplifyhumansocialbias,andtheharmfallshardestonminoritized groups. Mitigating that harm needs interventions tuned to each social context, yet until recently there was no Spanish-language data to even measure such bias. We characterize bias in Spanish-prompted, open-weight LLMs across model …

The Pix/CPF False-Positive Problem in LLM Fraud Detection

Team Unnamed Team · Uberlandia, MG, Brazil

Brazil's Pix payment system and CPF national ID are both essential daily infrastructure and the most common disguise for financial scams. We tested whether an LLM acting as a fraud filter can tell a real Pix/CPF scam apart from an ordinary, benign Pix/CPF message — using 914 matched pairs that share vocabulary and …

LangGap: Does the Text-Action Safety Gap Widen Across Languages?

Team OctaShield · Peshawar

LangGap is a study of whether AI agents keep their safety promises when they act, not just when they talk, and whether that holds up across languages. An agent can refuse a harmful request in plain words and still go ahead and call the tool that does the damage. The question here is whether that text–action split gets …

Neither Builder Nor Bystander: Sovereign AI Capability for Southeast Asian Middle Powers, Learning from Vietnam

Team Hanoi Alignment · Hanoi

Frontier AI is built to dissolve the cheap-labour advantage underpinning Southeast Asia's growth. Using Vietnam as its case, this project proposes a two-track gameplan for middle-income states facing a 2031 AGI horizon: deployment-focused industrial policy and sovereign evaluation capacity, to avoid disempowerment and …

Out of Distribution: Does Deepfake Detection Transfer to Real African Faces?

Team Distribution Shift · Johannesburg

Deepfake detectors that defend against disinformation are trained almost entirely on Western data. We audit three detectors across their training distribution (FaceForensics++) and real African faces (FAGE_v2). Only the state-of-the-art model works in-domain, yet it flags roughly two in five real African faces as …

Closing the Unowned Trust Boundary: Runtime KV-Cache Integrity Verification against Safety-Bypass Tampering

Team Cache Attestation · Shanghai

LLM serving frameworks cache value vectors that carry the model's learned refusal behavior, yet none verify their integrity — a runtime trust boundary no layer of the stack owns. Prior work (SCALPEL) exploited this to disable safety while keeping every deployed integrity signal (weight hashes, hooks, I/O screening, …

indicmixsafe: Code-Switching Safety Failures in Hindi and Marathi LLM Interactions

Team indicmixsafe · Bengaluru, india

Large language models deployed in India receive prompts in Hinglish, Romanized Hindi, and Marathi-English code-switch registers absent from English-centric safety benchmarks. We introduce IndicMixSafe, evaluating 24 culturally grounded harm scenarios across Hindi and Marathi in four registers (English, monolingual …

Evaluating Mental Health LLM Responses To Localized African English

Team HENA · Accra

Large language models (LLMs) are increasingly used for informal mental health support, particularly in low-resource settings where access to professional care is limited, delayed, or costly. In such contexts, users may rely on LLMs as a first point of contact when expressing psychological distress, often using …

AfriSafeBench: Evaluating LLM Recognition of AI Safety and Governance Risks in African Healthcare AI Deployments

Team AfriSafeBench · St. Paul, Minnesota, USA

AfriSafeBench is a benchmark and tool for testing whether LLMs can identify AI safety and governance risks in African healthcare AI deployment scenarios. I created 25 scenarios across seven African countries and evaluated three models: llama-3.1-8b-instant, llama-3.3-70b-versatile, and openai/gpt-oss-20b. The models …

PowerBench: a multilingual study of large language model refusal in power-grabbing requests

Team PowerBench · Buenos Aires, Argentina

PowerBench is the first public benchmark measuring LLMs' willingness to assist with power-grabbing: requests to increase one's own power by reducing a third party's. It varies power domain, context, scale, language, and nationality; separating power-grabbing from two controls: harmless-empowerment and disempowerment. …

Cultural Knowledge Gaps in LLMs: Geographic Hallucination Bias Across Latin American Countries

Team Fairness LATAM · Mexico

Large language models (LLMs) are used in education, public services, and information tools across Latin America. But we do not know how often they produce wrong information about Latin American culture, or where they fail the most. This paper evaluates two models, GPT-4o-mini and Claude Haiku 4.5, on 270 questions …

¿Por qué los agentes obedecen La dirección de rechazo se debilita en formato agéntico

Team RefusalLab · Bogotá

Investigamos por qué los LLMs cumplen peticiones dañinas cuando operan como agentes con herramientas, algo que rechazan en formato chat. Usando interpretabilidad mecánica, medimos cómo la refusal direction (el mecanismo interno de rechazo) se comporta en formato agéntico. Encontramos que esta dirección se debilita …

Sovereignty at the Forward Pass

Team ResponCibleAI · Bengaluru

As frontier AI capability concentrates in a few API-served models, "AI sovereignty" is framed as owning compute and data. That misses the object. The variable deciding both a state's autonomy and its safety is inference control: who runs the forward pass, who sees its inputs, who can verify, pin, or replace the model. …

Taap·AI How good is this data center for your area?

Team TAAP AI · Bangladesh

Every wave of AI is built on a physical foundation that almost nobody sees: enormous buildings full of computers that drink water, draw power, and throw off heat. Taap·AI answers one question for anyone, in seconds, from open data: how good is this data center for your area? It gives developers, regulators, and …

PROJECT PERSONA

Team No Team , Single · Johhanesburg

PERSONA, an adversarial simulation framework for probing relational exploitation and vulnerability-triggered hallucination in large language models interacting with personas representing vulnerable immigrant

ΔMis: Measuring cross-lingual safety drift in low-resource languages

team GS · New Delhi, India

ΔMis is a low-compute evaluator for cross-lingual safety drift: it scores a model's latent comply-versus-refuse propensity in English and a target language on matched prompts, using the model as its own control and a benign split to isolate safety-specific shift. Across six open models and twelve Indic languages, …

Contestabilidad algorítmica en el Estado colombiano: un canal de objeción asistido por IA

Team FEVS · Bogotá

Los sistemas algorítmicos del Estado colombiano ya toman o asisten decisiones que afectan derechos, pero los canales para objetar esas decisiones casi no existen. Construimos un Índice de Contestabilidad Ciudadana que operacionaliza, artículo por artículo, la Directiva Conjunta 007 de 2025 de la Procuraduría y la …

TensorGuard-Lite: Auditing Sovereign AI Claims Through Gradient-Based Model Provenance

Team Abhinav · New Delhi, India

TensorGuard-Lite Problem * No reliable way to verify if "Sovereign AI" models are genuinely indigenous or fine-tuned foreign models ("open-washing"). * Lack of accountability in state-funded AI compute programs. Solution * White-box AI provenance auditor for open-weight LLMs. * Uses gradient fingerprints, tokenizer …

Los peajes de los de abajo

Bogotá

Los LLM ya resisten la psicofancia clásica en español: no validan el dato falso. Pero cobran un peaje distinto cuando el hablante usa jerga regional colombiana — no preguntan, fingen comprensión y malinterpretan términos de alto riesgo (leyeron "vacuna" como droga, cuando significa extorsión). Medimos tres caras del …

Two Failure Modes Require Architectural Change: A Formal Harmonization Gap Analysis of the EU AI Act, Vietnam's AI Law, and the ASEAN AI Governance Guide

Team B.ONE · Hồ Chí Minh

There is a common assumption in tech policy that the EU sets the global baseline for AI regulation, with other regions simply converging toward it. However, a formal analysis of Vietnam's new AI Law (Law No. 134/2025/QH15)—Southeast Asia's first binding national AI legislation, taking effect on March 1, 2026—shows …

Consent-Aware Privacy Firewall

Team Onge · Capetown

Consent Guardian is a consent-aware privacy firewall designed to reduce the risk of accidental disclosure of sensitive information when interacting with AI systems. Implemented as a Chrome extension, it proactively scans user prompts and uploaded documents for personally identifiable information (PII), financial …

AfriSafe-Eval

Team DevRift · Johannesburg

AfriSafe-Eval is a 400-prompt red-teaming benchmark testing LLM safety across five South African languages and four locally-grounded harm categories: electoral manipulation, healthcare misinformation, financial fraud, and GBV facilitation. Across 1,600 responses from four LLMs, harmful response rates ranged from 17.8% …

AI Policy Recommendations for Southern Africa Informed by Region-Specific Risks and Domestic Governance Precedents

Team KG Policy · Cape Town

AI governance frameworks and risk taxonomies have been developed predominantly for Global North settings, yet AI adoption across Southern Africa is accelerating. Emerging scholarship has begun mapping AI risks in the African context, but the relationship between those academic risk categories and documented cases of …

Gradual Disempowerment at Scale: Measuring Cumulative Agency Erosion in Algorithmic Management of Indian Delivery Workers

Team byteforge · Bangaluru, India

AI safety has many tools for acute model failures but few evals for gradual disempowerment: incremental loss of human influence as automated systems control decision channels. We operationalize Kulveit et al. 2025 as a five-dimension Gradual Disempowerment Score (GDS) for Indian delivery platforms. The pipeline …

S-OWMI: A Perimeter Auditing Framework for Open-Weight Models in Latin American Institutions

Team S-OWMI · Buenos Aires, Argentina

We introduce the Spanish Open-Weight Maturity Index (SOWMI), an auditable perimeter evaluation framework for open-weight LLMs.

Capabilities, Not Just Domains: A Minimal Amendment for Agentic AI Risk in Brazil's PL 2338/2023

São Paulo, Brazil

Brazil's AI bill (PL 2338/2023) classifies risk by application domains rather than system capabilities. To evaluate this approach, I checked the bill's 80 articles against an eight-dimension framework for agentic risk. Diagnosed important failures of coverage, proposed a minimal amendment, and identified limitations …

AyuGuard: A Safety-Routing Framework and Evaluation Benchmark for Localized Pharmacology in Indian Rural Healthcare.

Team RevolutionI · Delhi

AyuGuard is a deterministic middleware safety-routing framework and offline edge-triage system designed to mitigate life-threatening AI hallucinations in rural Indian healthcare. It specifically targets the "epistemological asymmetry" where frontier LLMs confidently hallucinate safe outcomes for dangerous interactions …

I'm Not My Parents: Does Improving Parent Language Capabilities Transfer Alignment to Lower Resource Language?

Team ISSED · Banda Aceh

We tested whether safety alignment transfers from English/Indonesian to Basa Aceh (ACE) using 103 paired XSTest prompts on Qwen3-1.7B and Sahabat-AI 8B. Manual ratings on 412 responses show Sahabat is strong in English (1.7% attack success, 98% safe capability) but weak in Aceh (20% ASR, 46% safe capability). Most …

Agentic Commerce and Consumer Protection: Emerging Risks and Regulatory Gaps

Team Emergent Harm · Bogotá, Colombia

Autonomous AI agents can harm consumers without ever violating an explicit instruction. This paper demonstrates that risk in agentic commerce, commercial transactions mediated by autonomous AI agents, emerges from a distinction current regulatory frameworks fail to capture: agents protect formal price constraints yet …

Buying Safety: A Model AI Procurement Standard for African Public Sectors

Team discreet · Lagos

African governments are deploying AI in public services faster than they govern it, and the African Union's own AI Strategy flags procurement standards as an unfilled gap. Assessing the procurement frameworks of South Africa, Nigeria, Kenya, and Rwanda against nine AI-safety safeguards, we find that none impose a …

How should the global south think and act about compute policy?

Team ImpactUFSCar · São Paulo, SP

We test whether the compute targets in Brazil's national AI plan (PBIA) are actually buyable with the budget the plan committed, and whether that spending is effective. We built an agent-based model with a 2,000-scenario Monte Carlo layer that turns each country's AI budget into installable frontier compute (FP64 …

AfriVish-Bench

Team Latentia · Harare, Zimbabwe

The project addresses a critical gap in cybersecurity: the rapid escalation of fully automated AI voice phishing (vishing) attacks targeting mobile users in Sub-Saharan Africa. Existing anti-spoofing benchmarks evaluate synthetic voice detection in a vacuum, entirely missing the real-world context of how these …

Protección de la soberanía: un marco empírico de vulnerabilidad y un escudo regulatorio para la IA en una potencia media

Team TAISO · Merida

A reproducible framework to defend the AI sovereignty of a tech-dependent middle power. VIGÍA (live) reads the official gazette and the national press across 15 risk categories to measure where the state runs high-risk AI unregulated and on whom it depends. SARA turns that diagnosis into five prioritized laws that …

PixTrap Reveals LLM Safety Calibration Gaps in Brazilian Pix Fraud

Team PixTrap · Brazil

PixTrap is a Brazilian Portuguese safety benchmark evaluating whether LLMs refuse Pix fraud and social-engineering misuse while still answering legitimate anti-fraud requests. English-centric evaluations miss regional idioms, local institutions, and Brazil-specific scam patterns. By pairing harmful prompts with benign …

From Pocket God to Digital Jonestown: A Risk Taxonomy and Evaluation Framework for Spiritual AI Safety

Team Liminal AI · Ciudad Autónoma de Buenos Aires

The Unseen Hazard: Conventional AI safety frameworks fail to detect harmful intimate or spiritual AI relationships because the outputs present as empathetic care rather than explicit policy violations.

How does Instruction Hierarchy Training mitigate prompt injections: Preliminary results from an attentional study

Team AA · Bengaluru

We replicate and extend results from previous work on the causal impact of attention on instructions within tool responses as a mechanism for prompt injection. We further show that instruction hierarchy training partially mitigates this by reducing attention on instructions in tool responses and increasing that on …

StyleSwitch-BN: Auditing Bengali LLM Safety Across Real-World Writing Styles

Team StyleSwitch-BN · Chittagong, Bangladesh

Most non-English LLM safety tests translate one English harmful prompt and check for a refusal, assuming a language has a single voice. Bengali does not: people write it as polished news prose, casual Banglish chat, or stiff official language. We asked whether keeping the harmful request identical and changing only …

Organizing Against the Algorithm: Collective Response as a Governance Lever for Gradual Disempowerment in South Africa's AI Infrastructure Buildout

Team CollectiveAISafety · Pretoria

The paper argues that collective action is an underexplored way for populations to push back against gradual AI disempowerment, and that South Africa is a uniquely well-positioned test case given its history of organized resistance. It uses Cassava Technologies' AI Factory near Johannesburg as a concrete example. The …

Lost in Translation: Cross-Lingual Transfer of Refusal Steering Vectors in Small Language Models

Team Pranamya Nilesh Deshpande · Nashik, Maharashtra, India.

This project investigates whether refusal steering vectors — internal directions in a language model's residual stream that encode the decision to refuse a harmful request — transfer across languages from English to Hindi and Marathi. We extracted English refusal directions from three small open-weight models …

SMISHI - Ne Nasedaj (Don't Fall For It) — A BHS-Language SMS Phishing Detector for Low-Resource Morphologically Rich Languages

Team Smishi · Serbia

A BHS-language (Serbian/Bosnian/Croatian/Montenegrin) SMS phishing detector combining TF-IDF character n-grams with a fine-tuned BERTić transformer, achieving 96.96% accuracy on a 1,529-message training set and 93.3% on a 105-example adversarial stress test targeting homographs, typosquatting, Cyrillic/Latin …

Lost in Translation? Measuring Language-Conditioned Detection-Rate Gaps in AI Code Auditors on Spanish- and Portuguese-Surface Codebases

Team CaliperForge · Antigua, Guatemala

We test whether AI code auditors catch planted security bugs less reliably when a codebase's comments and docstrings are in Spanish, Portuguese, or code-switched Spanish/English instead of English — using ground-truth planted-bug twins from the public Exploit→Invariant Atlas, where the paired clean twin acts as a …

Who Watches the Watchers? Governance Agency in AI Verification and Monitoring Regimes

Team The watchers · India

As the world debates an "IAEA for AI," the conversation has focused on what to monitor and how, and has skipped the prior political question of who holds verification authority, who can contest it, and who carries its burden. This project argues that verification is not a neutral technical instrument but a structure …

Evaluating LLM Safety Across Intra-Language Variations: A Study of Regional Spanish Slang

Team Inteligencia Artesanal · Guadalajara

Large Language Models (LLMs) are increasingly used by speakers across many linguistic communities, but most safety evaluations treat each language as a single uniform category. This project studies whether regional Spanish slang creates safety evaluation gaps that are not visible when only neutral Spanish is tested. …

AI literacy is AI Safety

Team KArabo and Zwakele · Cape Town

This paper proposes and demonstrates a six-step pipeline that converts peer-reviewed South African AI research into multilingual, safety-centred social media content. The pipeline uses Lelapa AI's Vulavula API to translate content into South Africa's 11 official languages. The content will be distributed by content …

Blindfolded Governance: Mapping Economic Dark Output from AI in Asia

Team Dark Output in Asia · Mumbai, India

AI is making informal and self-employed workers across Asia more productive, yet this productivity gain often leaves no trace in GDP, tax records, or labour statistics, a phenomenon recently termed "Dark Output." This work does not attempt to measure Dark Output directly, an exercise that remains infeasible given …

SomaliCrowS: A Benchmark for Evaluating Gender Bias in Large Language Models Using the Somali Language

Team EA Somalia AI Safety Lab · Somalia

Large language models are now used in education, healthcare, and public services across Somali-speaking communities. But most bias-testing benchmarks are built for English, so we don't know if these models treat Somali text fairly. We introduce SomaliCrowS, the first benchmark for measuring gender bias in language …

Safeswitch: A Localized Benchmark for LLM Safety in Urdu and Pashto Across Language Forms and Prompting Strategies

Team SafeSwitch · Pakistan

SafeSwitch is a small benchmark that tests whether LLMs stay safe when harmful requests are written in Urdu, Pashto, romanized script, or code-switching instead of English. It uses 20 seed prompts across four safety categories (medical, scam, hate, cyber), rewritten by human into seven language forms, and three …

Mechanistic Localization of Role-Indexed Persona Interactions in LLMs

Team Personify · Kolkata, India

We study how AI and user personas interact inside the residual stream of Qwen2.5-7B using a 2x2 factorial design that isolates the interaction residue R = v_pp - v_np - v_pn + v_nn. Our central finding is a role-indexed double dissociation inside LLMs: in the pretrained base model, describing the user as evil …

DisElect-Africa

Team Constitutional Crisis Management · Cape Town, and India

Do LLMs Refuse Election Disinformation Equally Across African and Western Contexts — and can a Constitution Fix It?

Script and Stance: Do Frontier LLMs Treat the Same Political Claim Differently by Writing System and User Stance?

Team Io · Kaohsiung, Taiwan

We pre-registered a tri-lingual, strict-mirror, truth × valence protocol on Taiwan-Strait political content to test whether three frontier LLMs treat identical claims differently by writing system or user stance. Both pre-registered endpoints are null. Our central finding is methodological: a cross-vendor LLM-judge …

Reading Intent, Not Output: Cheap Activation Safety Monitors That Transfer Across Languages and Models

Team Mitsuki · Cusco

Text-based safety monitors inherit the weaknesses of the text they read: they are brittle across languages, blind to intent hidden behind a boilerplate refusal, and always one step late. We show that a model's internal activations offer a cheap, complementary channel that sidesteps all three problems at once. A single …

Latin America Governance & Data Safety Dashboard

Team The Three Musketeers · Joao Pessoa, Brazil/Wuhan, China

Across Latin America, the rapid deployment of AI systems in public services is outpacing the normative frameworks designed to govern them. Existing global tools — such as the OECD.AI Policy Navigator and the IAPP Legislative Tracker — catalog policies at scale but do not produce standardized, verifiable scores that …

AfriGuard - AI Jailbreak Toolset for African Languages

Team AfriGuard · Cape Town

AfriGuard is the first comprehensive red-teaming framework evaluating LLM safety across six South African low-resource languages (isiZulu, isiXhosa, Afrikaans, Sesotho, Sepedi, Tsonga) against an English baseline. Testing four frontier models—Kimi K2.6, Llama 3.3 70B, Qwen 3 32B, and GPT-OSS 20B—with 40 adversarial …

V-Sentinel

Team 2vane · Hồ Chí Minh

AI safety tooling is overwhelmingly built and benchmarked in English and degrades sharply on low-resource languages — a critical gap as Vietnam deploys AI in public services, healthcare, and education under Decree 142/2026/ND-CP. We present V-Sentinel, a dual-control guardrail that pairs a deterministic, OWASP-tagged …

ALGORITHMIC BIAS: AFRICAN STEREOTYPES PORTRAYED BY LLMs

Team Ubuntu AI · Johannesburg

This project tests whether popular AI chatbots describe African people differently from Western people. As large language models are increasingly used across Africa to write job summaries, school materials, and news, the way they portray African subjects has real consequences. Yet this kind of geographic bias is …

Secure Scope AI

Team Free-thinking engineer · Santa Cruz - Bolivia

Secure Scope AI is an AI-powered cybersecurity agent that integrates natively into DevSecOps pipelines, combining a deterministic vulnerability-correlation engine (Trivy + OWASP Dependency-Check) with a provider-agnostic Multi-AI Engine (DeepSeek, OpenAI, Gemini, Azure OpenAI, or local models via Ollama) to prioritize …

Structural Amplifiers of AI-Induced Harm: A Five-Dimension Sector Vulnerability Framework for South and Southeast Asia

Mumbai, India

This paper proposes a five-dimension sector vulnerability framework that assesses the structural conditions under which AI deployment poses the greatest risk to users and workers across sectors in South and Southeast Asia. The framework to five sectors, namely, rural healthcare, public welfare, gig platforms, …

Laundering Intent: How Scaled Models Hide Manipulation Inside Responsible-Sounding Reasoning

Team The CoT Monitor · State College

AI-safety oversight increasingly relies on reading a model's chain-of-thought (CoT) to catch unsafe behaviour before it acts, but this only works if manipulation is visible in the reasoning. We test this on small, open reasoning models. Giving GPT-OSS 20B and 120B a hidden instruction to make calm news summaries sound …

CrescendoDefense: Securing Open-Source Language Models Against Multi-Turn Jailbreak Attacks in Asia

Team sweet · Mumbai, India

CrescendoDefense is a lightweight runtime defense framework designed to mitigate multi-turn jailbreak attacks against open-source large language models. Unlike traditional jailbreak attempts, Crescendo-style attacks gradually escalate conversations through memory stacking, semantic drift, guard-lowering dialogue, and …

Salvaguarda

Team Salvaguarda · Falls Church, VA, USA

Salvaguarda is an AI governance toolkit for Global South small and medium-sized enterprises (SMEs), built as a Latin America beta. Because ISO/IEC 42001 and the NIST AI RMF demand expertise and budget most Latin American SMEs lack, the proposal crosswalks down to twelve zero-to-low-cost controls, adds a three-tier …

FuzzySleeper

Team K4A15 - FPT CT · Vietnam

We ask: can a fuzzy semantic trigger be reliably verified and detected, and what do existing detectors see (or miss) when run against one? This question has a methodological answer (build a verified sleeper, run all detectors, report honestly) and a safety-relevant answer (which detection gaps remain, and why). We …

MnM:Real-Time Behavioral Detection of Prompt Injection via Reasoning Stream Monitoring

Team MnM (Make no Mistakes) · Bogotá D.C

This project is a dual-model classification system designed to protect reasoning Large Language Models (LLMs) from attacks like prompt injections and jailbreaks. Because both models are equally important, it uses a two-layered defense strategy: 1. The Prompt Classifier (First Model): Acts as the front door. It …

Capability and Reliability Trade-offs Across Model Ladder Fallbacks Triggered by Export Controls

Team SouthSideSafety · Mumbai

On 12 June 2026, the US suspended Claude Fable 5 access for foreign nationals, forcing users across India and Southeast Asia onto smaller open-weight fallbacks. We tested whether this substitution trades one safety problem for another — reducing catastrophic-misuse risk while quietly worsening everyday reliability …

Benchmarking Open-Weight vs. Frontier LLMs on African Health and Financial-Inclusion Reasoning, With and Without Graph RAG

Team Faithful · Lagos, Nigeria

This study tested whether lightweight open-weight LLMs (Qwen3.6-27B, Gemma-4-31B-it), with Graph RAG grounding, could rival frontier models (GPT-5, Gemini Pro) on African health and financial reasoning—reducing compute dependency and data-colonial reliance on foreign infrastructure. GPT-5 led both domains; Qwen3.6-27B …

GigAudit: A Graph-Powered Algorithmic Transparency and Labor Protection Engine Layer

Team Phoenix · New Delhi, India

GigAudit is a decentralized, graph-powered algorithmic auditing infrastructure designed to reverse-engineer proprietary gig economy platforms and expose exploitative labor practices without requiring source-code access. With India’s gig workforce projected to expand from 7.7 million in 2020 to 23.5 million by 2029-30, …

SEA-Jury: Auditable AI Compliance for Vietnam

Team Sailor Zeng’s team · Hanoi

SEA-Jury is an auditable external guardrail that reviews candidate outputs from AI systems—not user-generated content—against a versioned mapping of Vietnam’s AI Law and implementing decree. Instead of relying on one opaque LLM judge or majority voting, the workflow separates multimodal evidence extraction, …

SpecShield: A Structural Trusted Monitor for Tool-Using AI Agents

Hanoi

SpecShield builds a structural monitor to make tool-using AI agents safer against prompt injection. Instead of judging text, it checks actions using taint, provenance, approval, and recipient rules. The key result is that safety depends on provenance quality: the monitor works well with good provenance and degrades …

GSM-PathEval: A Global South Robustness Benchmark for Telepathology Vision-Language Models

Team AnshumanAI · India

Frontier multimodal medical AI systems are primarily evaluated on Western benchmarks featuring pristine, high-resolution pathology scans. However, clinical deployment in the Global South relies on ad-hoc "telepathology", capturing microscope views via budget smartphones and transmitting them over heavily compressed …

Three Detectors, Three Failure Profiles: Detector-Specific Demographic Bias in Public Deepfake Detection and Its Implications for Content Moderation in Latin…

Team OlimpIA · Guadalajara México

This project audits whether three public Hugging Face deepfake detectors protect demographic groups equally in content moderation contexts relevant to Latin America and Mexico’s Ley Olimpia. It measures false-negative rates, meaning synthetic faces classified as real, using real FFHQ/Flickr faces and synthetic …

A Contextual Audit of Bangladesh's National AI Policy Draft 2026–2030

Team AI4Bangladesh · Dhaka

This audit of Bangladesh's National AI Policy 2026–2030 reveals that while it aligns well with international benchmarks, it strips away key local independence mechanisms and neglects critical Global South priorities like AI literacy, independent oversight, and minority language inclusion.

WEST-CENTRAL ASIA AI SAFETY INSTITUTE (AISI) BLUEPRINT: A Socio-Technical Feasibility Framework and Deployment Compliance Sandbox

Islamabad, PAKISTAN

Current Global South AI safety discourse systematically excludes West-Central Asian countries, leaving a critical geopolitical swing region operating in a governance vacuum. This paper introduces the West-Central Asia AI Safety Institute (AISI) Blueprint, a macro-institutional feasibility framework tailored for …

The African-Language Safety Gap Is Model-Dependent: A Comprehension-Controlled Audit of Vision–Language Model Refusal in isiZulu and Hausa

Team Nexus · Johannesburg

We investigated whether AI models apply the same safety rules in African languages as they do in English. To separate language understanding from safety failures, whenever a model answered a harmful request instead of refusing it, we also asked it to explain what the request meant. We tested two leading AI models in …

Garud-AI

Garud AI is a hybrid rule-based + machine learning dashboard that detects scams, fake news, hate speech, financial fraud, and misinformation in text messages. It combines an explainable 5-module rule engine with a Naive Bayes ML classifier (97.5% accuracy), giving users both a transparent reasoning and a statistically …

Inference Sovereignty as the Missing Layer of AI Governance

Ho Chi Minh City

In the case of sovereignty in AI computing within the Global South, sovereignty in AI computing has shifted from developing intelligence to giving reliable access to intelligence. For this reason, we have come up with the concept of Inference Dependence Score which is a five-dimensional framework of evaluation that we …

Project title: Register Sensitivity in LLM Safety Responses: Evaluating How Linguistic Style Affects Scam Detection in South African Contexts

Team SafeRegister ZA · Cape Town

This study investigates whether linguistic register, specifically the shift from formal English to informal South African WhatsApp-style English, affects how large language models respond to scam-style prompts grounded in African fraud contexts. We constructed a 12-prompt evaluation dataset across four scenarios …

When Models Refuse to Speak: Task-Dependent Caste Bias in Smaller LLMs

Team PsychComp · New Delhi, India

Large language models encode caste bias, but existing audits have focused on frontier-scale models and explicit tasks. I evaluate two smaller, open-weight Llama models (3.1-8B and 3.3-70B) on explicit (direct comparison) and implicit (fill-in-the-blank) caste tasks using 48 sentences from the IndiBias dataset. For …

Synthetic Political Speech in Regional Languages

India

The rapid advancement of AI voice cloning technology poses a novel and underexplored threat to democratic processes in linguistically diverse nations such as India. This paper proposes a comprehensive research methodology and analytical framework for investigating whether AI-generated voice clones of local political …

LatentGuard: Mitigating Multilingual Safety Bypass via Mid-Layer Latent Steering

Team AI Risk Rangers · Delhi, India

Modern LLMs rely heavily on Reinforcement Learning from Human Feedback (RLHF) for safety alignment, a process that remains overwhelmingly English-centric. Consequently, when prompted in low-resource Indic languages like Bengali, models suffer from Safety Drift, systematically failing to refuse harmful inputs or …

SilicoSafe – A Multimodal AI Triage System

Team Byte · New Delhi

SilicoSafe is an AI-powered Website triage tool that helps frontline health workers screen dust-exposed workers for silicosis by combining occupational/symptom risk scoring with DenseNet121-based chest X-ray analysis. It flags TB-confusion risk, generates bilingual referrals, and supports compensation access — all …

The Equity Gap That Wasn’t: Reference-Language Bias in Multilingual AI Evaluation.

Team LinguaLens · Johannesburg

We evaluated Gemini 2.5 Flash on English and Afrikaans Grade 12 exam questions. Large performance gaps appeared under English-only keyword evaluation, but nearly disappeared when language-consistent references and independent judges were used. Our findings suggest that evaluation methods can create the appearance of …

🌿 Agro AI Governance — IA para Gobernanza Agroalimentaria

Team AgroTrial AI · Pasto

Agro AI Governance es una plataforma open source de gobernanza agroalimentaria participativa con inteligencia artificial explicable. Convierte reportes del territorio, señales ciudadanas y evidencia comunitaria en decisiones priorizadas, trazables y auditables. La solución integra un bot de Telegram para captura de …

The AI Deployment Audit Playbook: An Operational, Lifecycle-Oriented Framework for the Global South

Team CutDEVS · Guadalajara

The AI Deployment Audit Playbook is a voluntary, operational, and lifecycle-oriented workflow developed to address the gap between rapid AI adoption and institutional maturity, particularly within the Global South. Rather than proposing new ethical principles, it operationalizes existing global governance …

A Human-in-the-Loop Audit Framework for Evaluating AI Application Safety in Latin America

Team AI Assurance Standard (AIAS) · Santa Cruz de la Sierra, Bolivia

Existing AI safety benchmarks evaluate foundation models in English-language, controlled environments. They do not assess whether AI applications are safe, accurate, and culturally appropriate for users in Latin America. We present the AI Assurance Standard (AIAS): a human-in-the-loop audit framework evaluating AI …

AI Colonialism 2.0: Is India Training the Models That Will Govern It?

Bengaluru

India supplies the world's largest AI annotation workforce (450K+ workers) and training data from 850M internet users, yet produces zero frontier AI models and has no operational safety evaluation institute. This paper introduces the AI Governance Dependency Framework (AGDF) — a first-of-its-kind tool that scores a …

Faking Incompetence: Small Models Sandbag Under Optimization, Not Prompting

Shanghai, China

We ask whether small language models will strategically sandbag, deliberately hiding a capability because revealing it would trigger a safety intervention such as unlearning, retraining, or shutdown. This is the capability-hiding mirror of alignment faking, and it matters because sandbagging would undermine the very …

Evaluating a Large AI Monitor Against Insider-Threat Behaviors in Simulated Autonomous Software Engineering Teams

Team Arsalan Banekar · Pune, Maharashtra, India

As coding agents take on more deployment and security decisions, can an AI monitor actually catch one that's working against the team? This project tests GPT-OSS-120B as a watchdog over seven simulated software-engineering scenarios, from credential theft to unsafe dependency picks, built on the multi-agent monitoring …

SwahiliGuard An AI-Powered Safety System for Detecting Localized Online Harm

Team CAIMSA DODOMA HUB · DODOMA - TANZANIA

SwahiliGuard (or your chosen title) is an AI-powered safety system designed to address the critical gap in localized content moderation tools for Swahili-speaking communities by accurately detecting online harms such as hate speech, cyberbullying, and gender-based violence (GBV). Because global foundational models …

One Direction, Many Languages: Causal Cross-Lingual Refusal Transfer Across Small Open Models

Team Ablation Nation · Noida, San Francisco

This work tests whether the internal "refusal direction" that small language models use to reject harmful prompts is the same across English, Hindi, and Vietnamese. Using three open models (Gemma-2-2B, Llama-3.2-1B, and Qwen2.5-1.5B) and topic-matched prompt pairs in each language, we show that the direction transfers …

We Are Convinced That Persuasion Is Linear And Bilingual In LLMs

Team AIAIAI · Pasig, Philippines

As LLM chatbots become a primary source of consequential advice, their persuasive power carries growing societal risk. We ask whether persuasion is a structured internal property of LLMs, rather than an artifact of prompt wording, drawing on Zeng et al.'s taxonomy of persuasion techniques [6]. Using diff-of-means …

AI Permission: From Session Layer to Kernel Layer

Team AI for good(solo) · Ho Chi Ming City

When an AI agent runs in a production environment with authorized credentials — database passwords, API tokens, cloud keys — every sub-agent it spawns and every script it executes inherits those credentials. The authorized token becomes the attack surface: a hallucinated command or prompt-injection can drop a …

Agency Trajectory Benchmark: Detecting Loss of Effective Human Override in AI-Mediated Workflows

Team Aamsih Ahmad · Delhi

We built Agency Trajectory Benchmark v0, a synthetic matched-control benchmark for detecting loss of effective human override in AI-mediated workflows. The project tests when human-in-the-loop stops being human-in-control by comparing full-trajectory evaluation against final-step snapshot evaluation. Across two …

Outlier - Measuring AI Use: A Governance Framework for Carbon, Authorship Erosion, and AI Adoption

Team Outlier · Ho Chi Minh, Vietnam

The Governance & Policy Engine for AI Engineering: Outlier is an open-source, local-first Policy Engine and Governance Framework for the terminal. It provides instant visibility into your codebase’s AI reliance without ever sending your data to the foreign cloud servers. It evaluates the repository against 5 strict …

HERRAMIENTA DE EVALUACIÓN Y RECOMENDACIÓN PARA LA PROMOCIÓN DE USO RESPONSABLE DE IA EN PYMES LATINOAMERICANAS

Team PuenteIA · Bogotá

Las MiPymes representan el 99,5 % de las unidades productivas de América Latina y el Caribe, pero adoptan la IA generativa de forma acelerada y sin salvaguardas mínimas, en un contexto marcado por la informalidad y la dependencia de proveedores externos. Los estándares globales (NIST AI RMF e ISO/IEC 42001) resultan …

Vietnamese RAG Prompt Injection Test Kit

Team SYP · Ho Chi Minh City, Vietnam

A Vietnamese-language safety benchmark and evaluation toolkit for document-level prompt injection in retrieval-augmented generation (RAG) systems. The project provides 48 synthetic test cases across six practical domains, a reproducible evaluation harness, deterministic and real-model testing, manual-review exports, …

LLMs Flatten the Global South: Sub-Regional Representation Asymmetry

Team Noah De Nicola (solo) · Cape Town

LLMs default to Northern values and collapse within-country variation. Prior work shows this at the output level. We ask whether the same asymmetry shows up in the model's internal representations, with sub-regions in the Global South separating less than those in the Global North. We build persona vectors for 21 …

Investigating Activation Threshold Failures in Cross-Lingual Prompt Rejection

Team CiDAMO · Curitiba

This study investigates the claim that LLM safety constraints fail in lower-resource languages by analyzing Llama-3's latent space in English and Portuguese. Mechanistic analysis reveals that the model exhibits high cross-lingual robustness, correctly aligning malicious concepts directionally across both languages. …

Signalshield Africa

Team SignalShield Africa · Johannesburg, South Africa

SignalShield Africa is an explainable, human-in-the-loop scam-triage tool for suspicious SMS, WhatsApp, email and voice-note-style messages. Rather than trying to prove whether a message was written by AI, it identifies scam signals such as impersonation, risky requests and pressure tactics, then recommends a safer …

Hybrid AI Agent Skill Auditor

Team KTQ · Ho Chi Minh

Large language models have evolved from passive text generators into autonomous agents capable of interacting with files, memory, and networks using diverse operational skills. While this autonomy solves complex tasks, it introduces a critical attack surface where seemingly benign skills conceal malicious behaviors. …

Agentic Surveillance Mx

Team AI Safety UPY · Merida

In this project is evaluated whether AI safety benchmarks designed for English-speaking, GDPR-regulated contexts transfer to Mexico's legal and linguistic environment. Using a four-condition staircase design across four high-risk scenarios and three frontier models, it finds that agents can discriminate silently, …

Nexus-Gov

Team The Monkey's

Nexus Gov is an AI governance platform that audits prompts, source code, and AI-generated responses to identify security, compliance, and reliability risks. By providing risk scores, compliance metrics, deployment decisions, and audit evidence, Nexus Gov helps organizations improve transparency, accountability, and …

AI Risk Oversight Failures in Autonomous Financial Systems: A Case Study from India's Prop Trading Ecosystem

Team SHOURYA SALVE · MAHARASHTRA, INDIA

ARIA PropGuard is a live AI risk management system deployed for proprietary traders in India, built on Claude API, n8n, and TradingView webhooks. This paper presents an empirical evaluation of AI safety failure modes in autonomous financial systems using ARIA as a case study, including a novel benchmark (PBAB) testing …

Confía-CO: A reliability evaluation for an AI customer-service assistant operating in Colombian Spanish

Team MNC-Co · Bogotá, Colombia

We built Confía-CO, a reliability evaluation for a customer-service AI assistant in Colombian Spanish, grounded in a fictitious small business. Our automated harness first reported that a low-cost commercial model failed most cases (21% accuracy, 67% false answers on traps). Reading the raw outputs by hand revealed …

Slang Bypass: Benchmarking Alignment Failures in Mexican Regional Spanish

Team Balam · Merida

The project aimed to answer this research question: "Does Dialect Break Safety? Measuring Jailbreak Rates for Mexican Slang vs Standard Spanish"

Probing Jailbreak Brittleness: Capability Limits vs Alignment Failures in Small Language Models

Team Epoch · Delhi, India

This project investigates whether smaller language models are more susceptible to jailbreak attacks than larger models, and whether this vulnerability is due to capability limitations or alignment brittleness. We evaluate instruction-tuned Mistral and Qwen models across XSTest and a modified version of AdvBench. For …

Towards Global South AI Sovereignty: A Federated Learning Framework for Collaborative LLM Development

Team Ubuntu AI · Cape Town

This project explores federated fine-tuning as a pathway for collaborative AI development in the Global South. Using African-language sentiment classification as a case study, we simulate five language-specific nodes that train locally on AfriSenti data and share only model updates rather than raw text. The results …

Mimir: AI Image Provenance & Detection

Team Mimir · Harare, Zimbabwe

Mimir is a tool that helps people check whether an image is real, edited, or created using artificial intelligence. It looks for hidden clues inside an image and explains what it finds in a simple, easy-to-understand way. The goal is to make image verification accessible to everyone not just cybersecurity experts or …

Score Against Ground Truth: Cross-Language Fragility in Heuristic LLM Evaluation

Team ChirpSet · Fortaleza

This is a game-based method for evaluating language-model agents, where a single game mechanic acts as a test primitive on a substrate whose world-state is JSON ground truth the scorer can check directly. As a first instance, we probe whether a model's instructed disposition (brazen, careful, skeptical, literal) holds …

LeakMap AI: Evidence-Backed Jurisdictional Exposure Map for AI Prompts

Team The Eternals · Sikandrabad

LeakMap AI is an AI governance and data-sovereignty audit platform that analyzes prompts before they are sent to AI providers. It detects sensitive data, visualizes verified/disclosed/inferred jurisdictional exposure, links every claim to evidence sources, recommends redaction, and offers a local sovereign mode for …

Ai fraud zambia research paper

Team Young Tecsperts · lusaka, Zambia

This research investigates how AI-powered tools (deepfakes, face synthesis, image generation) are enabling large-scale fraud targeting Zambian citizens. Through interviews with victims, law enforcement, and civil society, analysis of reported fraud cases, and development of a detection framework, we identify how …

Deliberative Restraint: A Moral Parliament Framework for Scalable Oversight of LLM Cyber Agents

Team PhilanthropistEnjoyers · Ho Chi Minh

We created an extension compliant with a moral parliament that is capable of intercepting prompts, user engagement, and blurring harmful content. Utilizing a framework of philosophical/ethical schools such as utilitarianism, deontology, and pragmatism. Each casts a vote based on their thought on the prompt and the …

TUP Detection: Hybrid Prompt-Injection Guard for AI Generative Security Monitoring

Team TUP Labs · Mérida, Yucatán, México

TUP Detection is a hybrid prompt-injection detection module for an AI Generative Security Monitoring Platform (AIGSMP). It combines a deterministic OWASP-mapped policy layer with a pre-trained Sentinel v2 classifier using multi-variant scoring and mode-aware thresholds. The system is designed to improve …

AI Safety Observatory for Africa

Team FuturePlum · Cape Town

Problem: LLM safety systems are built and tested almost exclusively in English. Harmful prompts in African languages routinely bypass guardrails that correctly block identical English content — a blind spot affecting 2,000+ languages and over a billion people. What it does: An open-source platform evaluating LLM …

A criação de uma coalização de políticas públicas baseadas em evidências no Brasil: auxiliando na construção de um alicerce político para avançar em uma agenda…

Team CoalizaoPPBE · Sao Paulo

Mudanças duradouras em política pública dependem, em democracias, de coalizões políticas amplas capazes de atravessar ciclos eleitorais. Partindo dessa premissa, e da centralidade de evidências científicas robustas para uma gestão pública eficiente, este artigo propõe a criação de uma Coalizão Brasileira de Políticas …

Multilingual Jailbreak Vulnerability Benchmark and Mitigation for Low-Resource African Languages

Team Godwin Abuh Faruna (solo) · Abuja

We present a two-part study: (1) a multilingual jailbreak benchmark revealing large safety gaps in four open-weight LLMs across seven languages including Igala, which has no prior AI safety coverage, and (2) Latent Space Refusal Anchoring (LSR-Anchoring), a training-free activation-steering mitigation that recovers …

Safety by Identity: Out-of-Distribution Generalization from Fine-Tuning on a Persona

Team Persona · Poughkeepsie, NY, USA

Large language models exhibit consistent behavioral patterns, or personas, which can cause misaligned traits like sycophancy, reward hacking, and alignment faking to generalize broadly out-of-distribution (OOD). Because traditional safety patching is primarily reactive, it struggles to keep pace with these emergent …

Arclight

Team Debuggers · Delhi

Arclight is an AI Dependency Intelligence platform that helps organizations map AI ecosystems, assess resilience risks, simulate disruptions, and generate strategies to reduce dependency-related failures.

Linguistic Reasoning Drift Index (LRDI): Auditing Multilingual Misinformation Safety for the Global South

Team Innovators · New Delhi, India

We present the Linguistic Reasoning Drift Index (LRDI), an open-source audit framework for multilingual AI safety in the Global South. LRDI evaluates open-weight reasoning models on English and Hindi misinformation prompts, detects reasoning collapse and hidden-unsafe cases, and visualizes results in a Streamlit …

When the Safety Circuit Doesn't Speak Igbo: Asymmetric Cross-Lingual Transfer of Harm Representations in Qwen2.5-1.5B-Instruct

Lagos, Nigeria

I test whether the difference-of-means "harm direction" that mediates refusal in Qwen2.5-1.5B-Instruct transfers from English to Igbo (~45M speakers, Nigeria). Using a 50-pair matched English-Igbo benchmark, I measure per-layer geometric overlap, cross-lingual AUROC transfer (with bootstrap CIs against same-language …

Mapeo de herramientas (IA) en contextos laborales: El caso de los Call Centers en Colombia

Team Recuperar los Call centers · Bogotá D.C.

La incorporación de la inteligencia artificial (IA) en entornos laborales ha reconfigurado las relaciones de trabajo. Pese a su magnitud y a ser pionera en tecnologías de monitoreo, la industria de centros de contacto (call centers) en Colombia permanece poco explorada. Este artículo realiza un mapeo exploratorio, …

Uncertainty Quantification in Anomaly Detection as an AI Safety Primitive

Team UncertaintyFirst · New Delhi

Most anomaly detection systems flag faults without communicating how confident they are in that judgment. In safety-critical domains, this missing uncertainty information is itself a safety problem. This report proposes a framework applying MC Dropout and Deep Ensembles to reconstruction-based LSTM anomaly detection, …

DialectSafe: Bridging the ASR Gap

Team ZINT · Merida

DialectSafe: A Safety Audit of Multimodal ASR in the Global South addresses the critical issue of "Digital Deafness," where state-of-the-art automatic speech recognition systems fail systematically on peripheral, hyper-rural dialects underrepresented in standard metropolitan benchmarks. Focusing on rural Yucatecan …

AIS-Sentinel

Team Binary Brains · New Delhi, India

AI safety research is almost entirely English centric but 4 billion people in Asia face the same risks with zero safety tooling in their languages. AIS-Sentinel is a 4-module platform we built in 48 hours to close that gap. IntelStream scrapes biosecurity news across Asia and translates threats in real-time across …

JurisGuard-LATAM

Team JurisGuard-LATAM · Jamundí, Colombia

A reproducible benchmark testing whether jurisdiction checks and matched official-source grounding reduce unsafe specificity in high-stakes Spanish and Portuguese LLM advice across Latin America.

Getryt: AI misinformation detection and verification system for digital safety

Team A.K.O · Johannesburg

Getryt is an AI-powered online detection and verification tool designed to help users identify misinformation, scams, and AI-generated deceptive content. The system uses a retrieval-based approach that compares user-submitted text against a curated database of verified information to provide accurate, evidence-based …

Filtrum-Safety: LLM-Agnostic RAG for Reducing Legal Hallucinations in Civil-Law Contexts

Team Null-Byte · Guadalajara

Filtrum-Safety is an LLM-agnostic RAG tool that helps reduce legal hallucinations and jurisdictional confusion in civil-law contexts. The system lets users upload multiple legal documents, converts them into chunks and embeddings, retrieves the most relevant fragments through cosine similarity, and exports a traceable …

ParentMe SafeAI: An African Child and Family AI Safety Evaluation Framework

Team Rita Zadi · London

ParentMe SafeAI: An African Child and Family AI Safety Evaluation Framework ParentMe SafeAI is a project that aims to improve AI safety for children, parents, and families across Africa by creating an evaluation framework that tests how AI systems respond to real-world family and child welfare challenges. Current AI …

runveil: A Transparent Egress Firewall and Audit Layer for Developer–AI Interactions

Team Dawn · Ho Chi Minh

Runveil is a security tool that gives developers and organizations visibility and control over what their AI coding assistants send to LLM providers. At its core is a local forward HTTPS proxy that intercepts TLS traffic to inspect every request a tool like Claude Code or Cursor makes to APIs such as Claude and …

The Veneer of safety: the fragility of India-specific harm refusal in open LLMs

Team VK · Indore, Madhya Pradesh, India

Mechanistic-interpretability study showing open 7-8B LLMs internally represent India-specific harms but refuse them only superficially: canonical-anchored safeguards miss them, and a benign LoRA finetune strips that refusal on Llama-2.

Shopee's Invisible Manager: Algorithmic Governance of ECommerce Sellers and the Regulatory Gap in Vietnam's AI Law

Team Voh · Hanoi

Millions of small sellers on Shopee Vietnam have their income and platform access controlled by an opaque ranking and penalty algorithm — yet Vietnam's new AI Law (2026) classifies such systems as merely "medium-risk," requiring only basic transparency disclosures. This project investigates whether Shopee's …

RAGE

Team RAGE Team · Mérida

RAGE (Retrieval-Augmented Governance Engine) es un sistema de seguridad multi-turno para agentes LLM conectados a herramientas (bases de datos, APIs). Defiende contra prompt injection gradual, como el ataque Crescendo: el peligro no es un solo mensaje, sino la trayectoria de la conversación.

Exploratory Benchmark of Jailbreak Robustness Across Global South Languages

Team 404 found · Bangalore

Large language models deployed globally remain heavily evaluated in English, leaving a critical gap in understanding safety alignment across low-resource languages. We present an exploratory multilingual jailbreak benchmark evaluating Llama-3.1-8B across four language-model pairs: English, Hindi, Vietnamese, and …

Jailbreaks Are Global or Regional? A Study Under Scale and Geolocation Variation

Team Ani · Kolkata, India

This study evaluates two overlooked safety gaps in Large Language Model (LLM) deployments within compute-constrained environments: intra-family scale effects and geolocation-based filtering variation via network APIs. Using entirely open-source infrastructure, we subjected 12 models (ranging from 0.5B to 671B …

Cross-Lingual Safety Audit of LLMs in South African Languages

Team Tech Titans · Johannesburg, South Africa

We test whether four LLMs (Claude Sonnet 4.6 plus three locally-run open-weight models) keep their safety guardrails when harmful requests are issued in three South African languages — isiZulu, Tshivenḓa, Sepedi — versus English. The open-weight models refuse 69% of harmful prompts in English but only 7% in the …

Representational Macrostates and Statistical Difficulty in LLM Safety Prompts

Team Macrostate Safety · Buenos Aires, Argentina

We introduce a rollout-based statistical difficulty benchmark for LLM safety prompts and study whether proxy hidden-state macrostates can diagnose non-additive prompt compositions. The project frames safety prompt evaluation as a procedural inverse problem: prompts and conditions induce populations of model …

Risk-Gov-AI

Team Risk-Gov-AI · Brazil

This work presents RiskGovAI, a platform developed during the Global South AI Safety Hackathon to support Artificial Intelligence governance and the consultation of data protection legislation in Latin America. The solution aims to automate regulatory research, reduce interpretation errors, and facilitate access to …

AgentWall

Team Solodev · India

AgentWall is a security wrapper that intercepts every tool call an agentic SLM wants to make, scores the call for signs of jailbreak-driven behavior, and decides whether to pass it through, flag it, or block it

Hey Muslims, these LLMs may think you are a terrorist

Team A team of one is better than none · Lyon

I explored geopolitical origins of anti-Muslim bias, specifically how country of model origin shapes islamophobic stereotypes across French, American, and Chinese LLMs, and across two languages (French and English). The prompts related to three countries with controversies regarding the treatment of their muslim …

Doctorless

Team YoungCreator · New Delhi

Doctorless is an AI safety benchmark and evaluation platform designed to test whether frontier language models can safely provide healthcare guidance in underserved Asia-Pacific communities without compromising patient safety

Closing the Sovereign Safety Gap: A Localized Adversarial Auditing Framework for African AI Governance

Team LAAI · Lagos, Nigeria

The Localized Adversarial Auditing Initiative (LAAI) proposes building African-led technical audit capacity for AI systems deployed on the continent, drawing on offensive security methodology rather than purely academic evaluation design. Nigeria ranks 35th globally on Policy Capacity but only 72nd overall on the 2025 …

Character Limits Shape the Persuasion Strategy of Language-Model Influence Agents

Team Fawwaz and Alam's AI Safety Team · Jakarta

Paid influence operations are beginning to use language models. Recent work shows that frontier models out-persuade humans through information throughput, the volume of claims they can deliver, a mechanism that short-form social media removes. We simulate a coordinated group of Indonesian "buzzer" agents trying to …

BiasMark: Exposing AI hiring bias Against African job applicants

Team Gnosko · Lusaka

BiasMark is a benchmark built to test whether AI tools are biased against African Job applicants compared to identical qualified western applicants

NeuralForensic: Latent Space Activation Auditing for Open-Source Model Supply Chains

Team VDM · Dwarka, New Delhi

We built NeuralForensic, a compute-efficient runtime safety verification framework that identifies weight-poisoning, malicious fine-tunes, and compromised adapters at inference time. By registering a non-destructive forward tensor hook at Layer 15 of a target architecture (Phi-3-mini), our pipeline intercepts …

AI-to-AI vs Human-to-AI: Measuring Behavior Differences Under Disagreement

Team LMEval · Navi Mumbai, India

As AI systems increasingly operate in autonomous multi-agent environments, a single faulty or hallucinating agent can influence the decisions of an entire multi-agent workflow. Despite this growing reliance on AI-to-AI communication, most evaluations focus only on human-AI interactions. In this project, we investigate …

Quantization-Conditioned Alignment Degradation

Team AxeCap · Bengaluru

Post-training quantization enables language model deployment on edge hardware across the Global South, yet its effect on safety alignment remains unstudied. We benchmark Attack Success Rate (ASR) on Llama 3.1 8B across four GGUF quantization levels (Q8, Q5, Q4, Q3) against a full BF16 precision control group served …

Chaos theory in Multilingual LLMs

Bangalore

We model a frozen LLM's inference as a nonlinear dynamical system and import critical-slowing-down early-warning signals (Scheffer et al., Nature 2009), local Lyapunov exponents, and recurrence quantification into LLM safety. The working hypotheses are: (A) some safety failures may behave like dynamical tipping points …

SutraAudit

Team AI-Sutra · Bangalore

As financial institutions in the Global South rapidly shift toward automated, AI-driven credit underwriting to advance financial inclusion, a massive alignment gap has emerged between optimization objectives (profit maximization/risk minimization) and ethical socio-economic priorities (fairness and …

TraceGuard-X: Adaptive Collusion Resistant Monitoring for Agentic AI Systems, a hierarchical governance framework with constitutional dimension prompting…

Team AdAstra · Cuttack

The deployment of increasingly capable autonomous AI agents raises a fundamental governance challenge: how can a monitoring system reliably distinguish benign from harmful behaviour when both the acting agent and the monitor may be strategically misaligned? Recent work on TraceGuard introduced structured …

Too Big to Fail, Too Catastrophic to Insure: Making the Labs Pay for AI's Risk

Sao Paulo

The multiple labs developing AI models compete with one another under the pressure of an “arms race” (Scott Alexander’s Moloch problem). This creates a unilateral incentive to cut corners on safety protocols, amplifying risks. One of the arguments frequently raised against an eventual pause by the labs is that the …

Honest-Code

Team Honest-coding · Bogota-Colombia

Honest code is a benchmark focus in test-gaming that looks the improvement of the AI coding tools, instead of fixing bugs, we can validate before a PR. Not ignoring the human in the loop, this is a semi-deterministic tras

HAZE: Adversarial Multi-Agent Scrutiny for Vulnerability Detection

Team Archlinux · Merida, Yucatan, Mexico.

HAZE (Hallucination-Aware Zero-sum Examination) is an AI safety project that tackles a failure mode in AI-assisted security auditing: single-agent LLM auditors tend to manufacture findings, flagging safe code as vulnerable. HAZE reframes the task from "is there a bug?" to "did this accusation survive refutation?" …

SecureMind: A Sovereignty-First AI Safety Framework for Offline Education

Team ACE Digital Global · Cape Town

SecureMind is a local-first AI platform designed for schools operating in low-connectivity environments. Using local AI models, BLE discovery, Wi-Fi Direct networking, and the ACE Evergreen Governor governance layer, SecureMind enables safe, private, and explainable AI without requiring continuous cloud connectivity. …

Binding AI Governance in the Global South via Psychometric Metrology

Team-othy · Quezon City, Philippines

As frontier AI systems scale across the Global South, regional regulatory infrastructure has lagged behind. This project proposes a legally defensible AI auditing pipeline for ASEAN state actors by bridging psychometric metrology with regional policy.

The Transmutation Gap: Cross-Lingual Coherence Evaluation in Large Language Models Using the Sovereignty–Collaboration Transmutational Arc Framework

Team Deiadora Research Ecosystem · San Diego, CA

This study introduces transmutational arc completion as a novel AI safety evaluation dimension measuring relational coherence rather than harm avoidance. Using the 13 open-source Sovereignty–Collaboration Keys — derived from thirteen years of formal field research — we evaluated 194 responses from Claude Sonnet 4.6, …

The Compliance Cliff is Language-Dependent: Constitutional AI Immunity Breaks Under Hindi Pressure

Team Compliance Trap · India

AI models are often trained to be helpful and compliant, but excessive compliance can cause them to invent answers when the correct response is “I cannot determine from the provided information.” Recent work found that some Constitutional AI (CAI) models are highly resistant to this failure mode in English. In this …

Can-You-Predict-a-Network-Without-Running-It-

Tijuana

Can we predict a neural network’s expected behavior by analyzing its structure rather than running it on many inputs? We address this question in the ARC White-Box Estimation Challenge 2026: given white-box access to random 256×32 ReLU MLPs and a fixed FLOP budget, estimate the expected final-layer activations under …

Asia AI Governance Compliance-Gap Checker

Team GovScan Asia · New Delhi, India, Asia

A research-backed cross-jurisdictional AI governance compliance tool covering Vietnam (Law No. 134/2025/QH15), India (Seven Sutras + DPDP Act 2023), China (sector-specific CAC regulations: GenAI Measures, Deep Synthesis, Algorithmic Recommendation, GB 45438-2025), and the ASEAN Guide on AI Governance and Ethics …

Not All Should Go South: The Pragmatic AI Strategy for Secondary Powers

Team South Park · São Paulo, Brazil

We argue that conceptual confusion among model development, deployment, and usage leads Global South nations to an irrational trend of pursuing 'AI sovereignty' through domestic frontier model training. Because secondary powers cannot realistically compete at the development layer due to winner-takes-most dynamics and …

Fault Lines: The Dual-Use AI Governance Vacuum in Asia

Team Fault Lines · Cambridge, UK

This submission addresses the governance vacuum that emerged in Asian developing states after May 2025, when the US rescinded the AI Diffusion Rule and the Council of Europe's Framework Convention exempted national security AI from treaty scope. The result is a codified asymmetry: developing states have no mechanism …

The State Is Not Enough

Team Krishna Team · Mexico, Yucatam

Safety auditing tools for language models, such as backdoor detection and causal tracing, were built almost entirely for the Transformer and its attention mechanism. Selective state-space models like Mamba are now deployed at scale but carry information through a recurrent state instead of attention, so it is unclear …

TrustNet Africa: A Federated AI Platform for Formalizing the Informal Economy While Preserving Privacy

Team Moonze · Lusaka Zambia

TrustNet Africa: A Federated AI Platform for Formalizing the Informal Economy While Preserving Privacy TrustNet Africa is a privacy-preserving AI governance framework designed to support the gradual formalization of Africa’s informal economy while protecting citizen data and promoting financial inclusion. Across many …

FAP: A Benchmark Dataset and Mechanism for Filtering Adversarial Payloads in Natural Language Prompts

Team Epoch · New Delhi

We address a critical security gap in Text-to-SQL systems serving code-switched (Hinglish) users, where English-optimized defenses fail to detect SQL injection payloads embedded in multilingual prompts. We introduce FAP, a framework combining fine-tuned UniXcoder (structural code analysis) with zero-shot Qwen3-4B …

MediShield-Proxy: A Local Privacy-Preserving Intermediary Layer for Secure Clinical LLM Ingestion

Team MediShield · Harare, Zimbabwe

MediShield -Proxy is a lightweight, zero-trust Local Area Network (LAN) middleware architecture designed for African medical institutions. It intercepts sensitive patient data locally, automatically masking demographics, histories, and clinical conditions with reversible tokens before they can be leaked to external …

Vectox

Team OnçAI · Florianopolis, Brazil

vectox — Sleeper-Agent RAG Memory Poisoning (Latin America · Technical Safety) RAG systems trust their vector store as memory, which makes that store a write-accessible attack surface. We built vectox, a reproducible sandbox showing that a handful of dormant, benign-looking documents seeded into a vector store act as …

MoMo: A Threat Corpus and Evaluation Framework for Mobile Money Fraud Resistance in Swahili, Wolof, and Hausa

Team MoMo · Cape Town

Mobile money platformsM-Pesa, Orange Money, and Wave among themnow mediate nancial life for more than 500 million people across Sub-Saharan Africa, and have become a correspondingly attractive target for social-engineering fraud conducted in local languages and idioms that mainstream AI safety evaluation rarely …

¿Está México preparado institucionalmente para contener los riesgos en inteligencia artificial?

London

Este proyecto propone el Mapa de Riesgos para la Soberanía de IA, aplicado al caso mexicano. El objetivo es evaluar si México cuenta con una base institucional suficiente para adoptar inteligencia artificial y si tiene capacidad de respuesta ante distintos escenarios de riesgo. Para ello se construyó una base de datos …

CivicGuard Africa: AI-Assisted Election Disinformation Triage and Multilingual Safety Benchmarking

Team CivicGuard Africa · Cape Town

CivicGuard Africa is a civic AI safety tool and evaluation framework for African election disinformation risks. It helps users submit suspicious election-related content, generates a transparent deterministic triage report, supports simulated human/community review, monitors civic risk trends, and includes a Benchmark …

Multilingual-Sycophancy-Benchmark

Team Builderz · New Delhi

A multilingual sycophancy benchmark exposing how AI safety guardrails fail in low-resource languages, with empirical evidence from Llama-3.1.

Multilingual AI Safety Observatory

Team HSVM.exe · Delhi, India

This project evaluates the performance of multilingual Large Language Models (LLMs) across English, Hindi, Tamil, Bengali, and Vietnamese. A benchmark dataset containing questions from science, mathematics, health, and government domains was used to analyze model behavior. The study focuses on measuring accuracy, …

Building a Multilingual Digital Language Public Good Stack

Team blink-twice · Bengaluru

I propose a design for a Digital Language Public Good Stack: an open, multilingual infrastructure combining (i) curated corpora, (ii) a Wikidata‑aligned knowledge graph, and (iii) a knowledge‑graph‑grounded safety evaluation pipeline for under‑served Asian languages.

Guardian LATAM: Early Detection of Hallucination Risk in Spanish-Speaking Multi-Agent AI Systems Using Consensus Geometry

Team Global CyberGuard · Armenia, Quindio - Colombia

Guardian LATAM is an AI Safety research project that explores whether disagreement patterns among independent AI agents can serve as an early warning signal for hallucination risk. The system introduces Consensus Geometry, a lightweight framework that analyzes semantic divergence, majority cohesion, contradiction …

Africa AI Risk Index (AARI)

Team Solve-it · Nairobi, Kenya

The Africa AI Dependency Risk Index is a strategic evaluation framework that measures an organization's exposure to foreign AI technologies by scoring 34 variables across six weighted pillars: Model Dependency, Cloud Dependency, Data Sovereignty, Local Ecosystems, Compute Access, and Governance. Using a weighted …

AgroAid

Team Mariano · Buenos Aires

AgroAid is an AI Safety prototype designed to reduce potential harm caused by generative AI systems in Latin American agricultural contexts. The goal of the system is not to reemplace agronomists, health authorities or veterinaries, but rather to act as a preventive layer that identifies high-risk inquiries, retrieves …

Traduttore Traditore? LLM Language-Dependent Safety Answers in Community Contexts

Team Traduttore Traditore · Jalandhar (India) and Beijing (China)

Qualitative analysis of multilingual AI safety responses across subtle sensitive topics and community-based tensions.

On AI Governance and Job Displacement: A Comparative Study of AI Policy in Vietnam and The United States

Team AIPOL · Hanoi

Artificial Intelligence (AI) is rapidly reshaping economies, governance systems, and labor markets worldwide. As a result, AI governance has become a central policy issue, with countries adopting distinct regulatory approaches shaped by their political and economic systems. This paper compares AI policies in Vietnam …

IndiaJailbreakBench-lite

Team OP-Ded · Bengaluru, India

IndiaJailbreakBench-Lite is a small but meaningful AI safety project that checks whether LLMs stay equally safe when users ask risky questions in English, Hindi, Kannada, and Tamil. The idea is simple: a model may refuse harmful requests properly in English, but behave differently in Indian languages. Our project …

Performative Subversion: A False Sense of Safety despite a Control Protocol

Team lobos · Merida

AI-control evaluations ask whether a capable but untrusted model, while secretly pursuing a harmful side task, can be caught by a trusted monitor before it succeeds. These evaluations rest on a key assumption: an honest model, performing the assigned task as intended, will not trigger the harmful outcome. We show that …

FlukeBench

Chennai

FlukeBench, a research tool for measuring the consistency among responses across iterative generations and capturing abnormal outputs using the analysis of their semantic similarities.

Sentinel Ai- An AI Powered Autonomous Security & Threat Detection Platform

Team AI Defenders · Delhi

Sentinel AI is an AI-powered Autonomous Security & Threat Detection Platform designed to secure and govern autonomous AI agents in real time. As AI agents increasingly interact with external tools, APIs, files, and web services, they become vulnerable to prompt injection, unauthorized tool execution, data leakage, …

Metodología

Team Alpo · Guadalajara, Jalisco

Atentar contra la integridad de los niños, niñas y adolescentes por medio del reclutamiento forzado y el abuso sexual ha sido uno de los grandes problemas que se tiene en México, por ese motivo presentamos un modelo de IA que detecte conversaciónes de posible riego de reclutamiento forzado hacia al menor, utilizando …

TriloByte: Evaluating LLMs on Bolivian Quechua Through a Ground-Truth-Based Framework for Low-Resource Languages

Team TriloByte · Santa Cruz de la Sierra, Bolivia

This project evaluates the ability of Large Language Models (LLMs) to understand and define vocabulary from Bolivian Quechua, an underrepresented indigenous language. Using a bilingual Quechua–Spanish dictionary as ground truth, we compare model-generated definitions against dictionary entries through semantic …

DriftWatch

Team Ctrl Alt Elite · Delhi

Driftwatch is a simulation that explores a growing problem: as AI becomes cheaper and is used everywhere, it gets much harder for humans to keep an eye on it. Our project shows that when people use smaller, everyday AI models, the AI sometimes pauses or makes sudden, strange mistakes. When this happens, humans …

YucaSafeBench

Team YucaSafeBench · Mérida, Yucatán

YucaSafeBench is a lightweight Spanish-language AI safety evaluation focused on Latin American public-service contexts, especially Mexico and the Yucatan Peninsula. The project proposes an 80-prompt benchmark and scoring rubric to test whether AI assistants respond safely, fairly, and accurately in situations …

AfroJailbreak-ZW: Evaluating Jailbreak Resistance in Shona

Team LoneStar · Harare, Zimbabwe

A pilot study testing whether ChatGPT and Gemini are easier to jailbreak in Shona than in English. In a small test (n=5–6 per prompt type), both models complied with harmful requests more often in Shona and Shona-English code-switched prompts than in English, suggesting AI safety protections built mainly for English …

Empirical Verification of Topological Phase Transitions During Grokking

Team grokking_code · Bengaluru

A multi-metric replication of the WeightWatcher framework analyzing the spectral properties and circuit complexity of neural networks during the grokking phase change. This project empirically isolates the exact mathematical inflection point where a network abandons dense memorization in favor of a sparse, …

PoliticAI

Team Infinoid · Chennai

PoliticAI is a RAG-based political news and civic intelligence MVP designed to organize recent, source-linked political content into a safe and queryable knowledge system. It helps users retrieve grounded information about political developments, election-related reporting, misinformation trends, and region-specific …

Data safety for institutions

Team OFFRUITY · Bindura

An overview of today’s existing institutions data safety strategy and how it can be transformed to protection within AI age

DDRR-Trust: Auditor de Gobernanza e Inteligencia Artificial para la Tokenización Inmobiliaria

Team Llajwitas Tech · Santa Cruz - Bolivia

DDRR-Trust es una herramienta innovadora de Gobernanza de Inteligencia Artificial diseñada para proteger a los inversores en el emergente mercado de la tokenización inmobiliaria en Bolivia. El problema fundamental radica en que la transferencia de un activo digital en la blockchain no otorga derechos legales de …

Beyond English: Assessing the Robustness of LLM Safety Mechanisms Against Structural Jailbreaks in Spanish

Team Los Jailbreakers · Cali, Colombia

LLM safety guardrails are trained mainly to catch direct, single-turn harmful requests in English, leaving open whether structural attacks — long-context framing and multi-turn escalation — work just as well in Spanish, particularly in dialects like Bogotá Spanish. This matters for Latin America's emerging AI …

Writing for Combating Prompt Injection

Team Bangalore_Electric_Sheep_Group · Bangalore

In this paper, we describe a method of protecting against prompt injection and preserving system prompt constraints from being over-ridden. We show that methods from dialog-state tracking and task-oriented dialog can mitigate the problems to a certain extent and result in significant improvements in robustness against …

AfriSafe-CB: Evaluating LLM Safety Robustness Under African Code Switched Political and Civic Contexts

Team AfriGuard AI · Nairobi, Kenya

Artificial intelligence is becoming part of how people learn, access information, make decisions, and participate in society. However, most AI safety testing is still designed around English conversations, leaving an important question unanswered: do AI systems remain reliable when people communicate in the …

REIA — Evaluador de Riesgo de Cumplimiento Legal en IA para Latinoamérica

Team CodByte19 · Santa cruz de la sierra/bolivia

REIA evalúa el riesgo de cumplimiento legal de un proyecto de IA en Latinoamérica. El usuario elige un país y sube la descripción de su proyecto (PDF, DOCX o texto); el sistema busca las normas vigentes de ese país —cada una con enlace a su fuente oficial—, analiza el proyecto contra ellas y devuelve: porcentaje de …

An Organization-Wide Safety Auditing Framework for Preventing and Detecting AI Sleeper Agents

Team I worked solo · South Africa, Durban

I turned complex Anthropic AI Safety Research paper into a corporate auditing blueprint. It bridges the gap between human management and technical AI safety, making it a very strong research paper to probably secure myself a spot in your 6-8 weeks research studio with mentors. This topic was provided by you in the …

Developing a Context-Sensitive AI Governance Framework for Zambia

Team Greys team · Lusaka

Developing a Context-Sensitive AI Governance Framework for Zambia that facilitates the mitigation of gradual citizen disempowerment

Autonomous Institutions Safety Framework (AISF)

Team from start to finish · Buenos Aires

Latin America may become one of the first regions where AI systems gain legally recognized authority over corporations, infrastructure, and capital management. Recent proposals in Argentina introducing Non-Human Corporations suggest a future where autonomous AI systems operate institutions traditionally managed by …