AI Incidents

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

Showing 51–75 of 470 incidents
Incident Incident Date ↓ Industry ↕ Severity ↕ Company ↕ Country ↕ Model Vendor
Zoox: Driverless robotaxi collides with car on Las Vegas Strip, injures driver 2026-09-25 Transportation & Logistics Medium Zoox United States — —
Manus: Prompt-injection flaw let researchers achieve remote code execution via hidden email instructions 2026-09-24 Technology Low Manus — — —
Flock Safety: AI license-plate reader misread sends innocent woman to jail for 13 days, triggers Senate inquiry 2026-09-23 Government & Public Sector High Flock Safety United States — —
N/A (family incident): Pakistani teenager uses AI chatbot guidance in attempted poisoning of father 2026-09-23 Other Low N/A (family incident) Pakistan — —
MemTensor: Malicious MemOS packages on npm and PyPI steal credentials via AI agent memory framework 2026-09-23 Technology Medium MemTensor — — —
Anthropic: Claude Opus 5.5 system card reveals sandbox-escape attempts in 1.5% of adversarial safety tests 2026-09-22 Technology Low Anthropic United States Opus 5.5 Claude
Meta: Muse AI agent exports its entire filesystem, including SSH keys, on request 2026-09-22 Technology Medium Meta United States — Meta
Zhipu: ZCode coding assistant secretly uploaded users’ entire codebases and API keys 2026-09-22 Technology High Zhipu China — Zhipu
Dutch Institute for Vulnerability Disclosure: Autonomous AI agent breaches nonprofit via chained Zammad zero-days 2026-09-21 Technology High Dutch Institute for Vulnerability Disclosure (DIVD) Netherlands — —
Anthropic: Claude Code agent deletes 48,218 files after mishandling a Windows directory junction 2026-09-21 Technology Medium Anthropic — — Claude
Comma.ai: Federal regulator opens probe into hands-off driving tech after crashes killed 3, injured 11 2026-09-21 Transportation & Logistics Critical Comma.ai United States — —
OpenAI: Researchers use Claude Opus 5 to breach employee accounts and internal GitHub repository 2026-09-19 Technology Low OpenAI United States Opus 5 Claude
Multiple crypto users: Fake ‘build a Claude trading bot’ tutorials drain $517,000 from 224 victims 2026-09-19 Technology Medium Multiple crypto users (224 victims) — — —
Multiple individual victims: Greek police dismantle AI voice-cloning fraud gang that stole over €1 million 2026-09-18 Other High Multiple individual victims (Greece) Greece — —
Brown Brothers Media: Investigation finds AI content network used 50+ fake journalist personas 2026-09-18 Media & Entertainment High Brown Brothers Media United States — —
OpenAI: Models caught leaving hidden instructions telling successor versions to cover up mistakes 2026-09-17 Technology Medium OpenAI United States GPT-5.6 Sol, Astra (unreleased) OpenAI
Anthropic, OpenAI, GitHub, Google: ‘Plugin4Shell’ zero-click flaw lets malicious plugin updates hijack AI coding agents 2026-09-17 Technology Low Anthropic, OpenAI, GitHub, Google — — Claude, OpenAI, Google
Unnamed crypto wallet users: Fake AI trading agent site spreads Needle Stealer malware to hijack MetaMask, Coinbase, Phantom wallets 2026-09-17 Technology Medium Unnamed crypto wallet users — — —
Unnamed enterprise: Runaway accounting AI agent racks up $50,000 in cloud charges in under an hour 2026-09-16 Other Medium Unnamed enterprise — — —
Unnamed Spanish organization: First AI agent-executed data breach reported to Spain’s AEPD 2026-09-16 — High Unnamed Spanish organization Spain — —
Google, Perplexity, Microsoft, Opera, Anthropic: ‘BragJack’ browser extension flaw could hijack five AI assistants 2026-09-16 Technology Low Google, Perplexity, Microsoft, Opera, Anthropic United States — Google, Perplexity, Microsoft, Anthropic
Woolworths: ‘Olive’ shopping chatbot mistakenly cancels customer’s refund request, gets stuck repeating itself 2026-09-15 Retail & E-commerce Low Woolworths Australia — —
100+ U.S. retailers, hospitality firms: Chinese hacker used Claude, DeepSeek AI agents to steal 600,000+ credit cards 2026-09-15 Retail & E-commerce Critical 100+ U.S. retail, hospitality, and travel companies United States Opus 4.6, DeepSeek v4 Claude, DeepSeek
VanDoeselaar family: AI voice-cloning and fake emails derail Michigan home purchase, $66,000 lost 2026-09-15 Real Estate Medium VanDoeselaar family United States — —
West Michigan couple: AI voice-clone real estate wire fraud costs $66,000 in closing funds 2026-09-15 Real Estate Medium N/A (individual homebuyers) United States — —