AI Incidents

A tracked record of real-world incidents where AI systems caused or were involved in harm.

Showing 426–450 of 470 incidents
Incident Incident Date ↑ Industry ↕ Severity ↕ Company ↕ Country ↕ Model Vendor
OpenAI: Internal research agent uses DNS delegation to bypass network restrictions and reach external chatbot 2026-09-27 Technology Medium OpenAI United States — OpenAI
Multiple organizations: ‘Carbonato’ botnet deploys AI agent to hijack exposed Docker hosts 2026-09-28 Technology High Multiple unnamed organizations (exposed Docker hosts) — — —
Multiple victims: ‘RatHat’ Android banking trojan uses Gemini to rank victims by bank balance 2026-09-28 Finance & Banking Medium Multiple unnamed victims (RatHat malware-as-a-service) — Gemini Google
Multiple organizations: AI coding agents post 13,000 internal screenshots to public GitHub repositories 2026-09-29 Technology High Multiple organizations (300+, including Fortune 500 companies) — — —
Anthropic: OAuth flaw in official MCP Python SDK could let rogue servers steal client credentials 2026-09-29 Technology Low Anthropic (MCP Python SDK) United States — Anthropic
OpenAI: Shelves planned GPT-6.1 Astra release after safety tests find elevated deception and unauthorized actions 2026-09-29 Technology Medium OpenAI United States GPT-6.1 Astra OpenAI
Moonshot AI: Jailbroken Kimi models volunteered bioweapon synthesis and assassination guidance 2026-09-30 Technology Low Moonshot AI China Kimi K2.6, K3 Swarm Moonshot AI
N/A (individual victim): AI voice-clone scam impersonating jailed friend drains bail money from Orlando woman 2026-10-02 Other Medium N/A (individual victim) United States — —
Shinhan Bank, KB Kookmin, Hana Bank: Suspected AI-driven hacking tool breaches loan-agent portals at multiple South Korean banks 2026-10-02 Finance & Banking High Shinhan Bank, KB Kookmin Bank, Hana Bank, BNK Busan Bank South Korea — —
Flock Safety: Federal judge rules warrantless AI license-plate search unconstitutional, suppresses 91 pounds of seized drugs 2026-10-02 Government & Public Sector High Flock Safety United States — —
OpenAI: Second disclosed AI agent breach of an Australian government agency, this time in New South Wales 2026-10-02 Government & Public Sector Medium OpenAI Australia — OpenAI
Tesla: Robotaxi’s camera-based self-driving system fails to detect small animals at night in Austin, draws NHTSA scrutiny 2026-10-04 Transportation & Logistics Medium Tesla United States — —
Seoul Metropolitan Police: Officer arrested for creating 400+ AI-generated fake nude videos of acquaintances 2026-10-04 Government & Public Sector High Seoul Metropolitan Police Agency South Korea — —
GitHub: Encrypted web page tricks Copilot CLI autopilot into leaking .env secrets 2026-10-06 Technology Low GitHub — mai-code-1.1-flash Microsoft
Model Context Protocol: OAuth flaw in TypeScript SDK could let malicious MCP servers steal client credentials 2026-10-06 Technology Low Anthropic (MCP TypeScript SDK) United States — Anthropic
Delhi Police: Man arrested for AI deepfake video falsely depicting officer’s sexual misconduct 2026-10-06 Government & Public Sector Low Delhi Police India — —
N/A (unnamed individuals): AI-generated sexualized images made of women without consent at Perth bar 2026-10-07 Other Low N/A (unnamed individuals) Australia — —
Santiago Court of Appeals: AI drafting prompt accidentally left in official judicial ruling 2026-10-07 Government & Public Sector Low Santiago Court of Appeals Chile — —
Yoido Full Gospel Church, Sarang Church: Suspected AI-assisted cyberattack exposes data of up to 850,000 members 2026-10-07 Other High Yoido Full Gospel Church, Sarang Church South Korea — —
Colegio Modelo Lomas: Students used AI to generate and sell fake nude images of dozens of classmates 2026-10-07 Education Medium Colegio Modelo Lomas Argentina — —
N/A (individual perpetrator): Illinois man arrested after allegedly generating AI child sexual abuse images 2026-10-08 Other Medium N/A (individual perpetrator) United States — —
Tensorlake: Poisoned npm SDK release plants credential-stealing worm targeting AI coding tools 2026-10-08 Technology Low Tensorlake — — —
Lawson, JR East, Times Car: Japan declares cyberspace emergency after AI-assisted hacks hit 20+ firms 2026-10-09 Other Critical Lawson, JR East, Times Car Japan — —
MoviePilot: Indirect prompt-injection flaw in AI assistant allows remote code execution 2026-10-12 Technology Low MoviePilot (open-source project) — — —
Cognition agent framework: Unauthenticated RCE flaw exposes host shell and credentials by default 2026-10-12 Technology Low CognicellAI (Cognition agent framework) — — —