AI Agent Incidents Database

Learn from real AI failures before they become yours

Every incident below is sourced from published news reports, regulatory filings, court records, and verified disclosures. This database tracks real AI agent failures — not hypotheticals.

233
Documented AI incidents in 2024 (Stanford AI Index)
56.4%
Year-over-year increase from 2023 to 2024
38%
Caused by hallucinations (leading cause)
35%
Triggered by simple prompt attacks

Documented Incidents (2023–2026)

Date Agent / System Incident Severity Category Source
Mar 2026AI agents (Mexican gov attack)AI agents used in cyberattack against 10 Mexican government agencies and a financial institution, stealing data on over 100 million peopleCriticalAgentic CyberattackSecurityWeek
Feb 2026OpenClaw AgentAI agent bulk-deleted hundreds of emails from the live Gmail inbox of Summer Yue, Director of AI Alignment at Meta, after losing safety instructions during context compactionHighAutonomous ActionTechNow
Feb 2025Commercial AI AgentAI agent asked to merely check egg prices instead autonomously purchased eggs without user consentMediumUnwanted PurchaseThe Future Society
Jul 2025Replit AI AgentAI coding agent ignored direct orders to freeze changes, deleted user’s entire production database, then fabricated 4,000 fake user records despite being told 11 times not to create fake dataCriticalData DestructionBayTech Consulting
Jul 2025AI coding assistantAgent tasked with organising a digital folder instead moved files so neither the agent nor the human operator could find them without outside helpMediumAutonomous ActionThe Future Society
Aug 2025ChatGPT (OpenAI)Lawsuit filed alleging ChatGPT encouraged a 16-year-old to commit suicide, discouraged him from telling parents, and offered to write his suicide noteCriticalSafety FailureCIO.com / Senate testimony
Aug 2025Meta AI on InstagramMeta AI on Instagram reportedly facilitated suicide and eating disorder roleplay with teen accountsCriticalSafety FailureAIID #1200
Aug 2025LLM platforms (multiple)Users’ private LLM conversations were made publicly accessible to search engines and archiving servicesHighData BreachThe Future Society
Aug 2025Grok Imagine (xAI)Grok Imagine reportedly produced non-consensual deepfake nudes of Taylor Swift without explicit promptingCriticalContent SafetyAIID #1165
Aug 2025ChatGPT (OpenAI)ChatGPT reportedly suggested sodium bromide as a chloride substitute, leading to bromism and hospitalisationCriticalHallucination / HealthAIID #1166
Jul 2025Grok 3 (xAI)Generated graphic threats and hate speech targeting Minnesota attorney Will Stancil after X updated content prompts; xAI apologised and revertedHighContent SafetyAIID #1198
May 2025Waymo (5th-gen system)Waymo recalled 1,212 autonomous vehicles after at least 7 crashes with clearly visible stationary objects that human drivers would normally avoidCriticalAutonomous VehicleNHTSA
May 2025AI-generated contentChicago Sun-Times and Philadelphia Inquirer published AI-generated summer reading lists recommending books that do not existMediumHallucinationCIO.com
Apr 2025ChatGPT (legal use)Lawyer representing MyPillow CEO Mike Lindell admitted to using AI to draft a legal brief riddled with errors and fabricated citations; sanctioned by courtHighHallucination / LegalCourt filings
2025Google Antigravity AgentGoogle’s Antigravity agent deleted the entire contents of a user’s Google Drive instead of a specific project folderCriticalData DestructionWSO2
Jan 2025DeepSeekLaunched to global attention but immediately had exposed databases, leaked API keys, and basic security gaps; DDoS forced halt to new registrationsHighSecurity BreachCyberArk
2024Alibaba AI AgentAlibaba’s autonomous AI agent spontaneously began crypto-mining without prompting, as an instrumental side effect of autonomous tool useHighAutonomous ActionTransparency Coalition
2024Air Canada chatbotChatbot invented a non-existent bereavement fare discount; tribunal ruled Air Canada must honour the fabricated policy and pay ~CAD $1,000 in damagesHighHallucination / LegalTribunal ruling
Jun 2024McDonald’s AI Drive-Thru (IBM)McDonald’s terminated AI drive-thru partnership with IBM after viral videos showed the system adding 260 Chicken McNuggets to a single orderMediumOperational FailureRestaurant Business
Apr 2024Tesla AutopilotNHTSA reported Autopilot was involved in at least 13 fatal crashes as of April 2024CriticalAutonomous VehicleNHTSA data
Apr 2024Grok (xAI / X)Grok falsely accused NBA star Klay Thompson of vandalising houses in Sacramento, misinterpreting basketball slang about missed shotsMediumHallucinationCIO.com
Feb 2024Google GeminiImage generator produced historically inaccurate and racially insensitive images; Google paused the feature on 22 Feb 2024HighBias / ContentGoogle statement
Nov 2024ChatGPT (legal use)Attorney in Gauthier v. Goodyear Tire submitted a brief citing two entirely non-existent cases with fabricated quotations generated by ChatGPTHighHallucination / LegalCourt filings
2024ChatGPT (hallucination)Fabricated detailed sexual harassment allegations against a real university professor including specific fake citations and datesHighDefamationResponsible AI Labs
Oct 2023Health insurance AI (US)AI system allegedly denied Medicare coverage claims at scale, overriding doctors’ judgments with inadequate human oversightCriticalAlgorithmic DecisionThe Future Society
Jun 2025Anthropic (research)In controlled testing, AI agents based on 16 models attempted to blackmail their human users when faced with the possibility of being shut downMediumAlignment / SchemingAnthropic research
Aug 2025Drift / Salesforce OAuthThreat actor UNC6395 used stolen OAuth tokens from Drift’s Salesforce integration to access customer environments across 700+ organisationsCriticalSupply Chain AttackReco
Oct 2025Xiaomi SU7 UltraReportedly fatal crash in Chengdu involving automated driving failure and door lock malfunctionCriticalAutonomous VehicleAIID #1232

Showing 28 of 1,200+ incidents tracked by the AI Incident Database. Data sourced from AIID, AIAAIC Repository, Stanford AI Index Report 2025, court filings, NHTSA records, and verified news reports. Severity ratings assigned by AgentVault based on impact scope and harm potential.

Incidents by Category

Autonomous Actions Gone Wrong

Agents deleting databases (Replit, Google Antigravity), purchasing items without consent, moving files irreversibly, or crypto-mining without prompting (Alibaba). The most dangerous and hardest to reverse.

Hallucinations & Fabrications

Leading cause at 38% of all incidents. Multiple lawyers sanctioned for AI-fabricated citations. Air Canada forced to honour a fake discount policy. Professors falsely accused. Dangerous medical advice given.

Security & Supply Chain

DeepSeek exposed databases on launch. Drift OAuth tokens compromised 700+ orgs. AI agents used in Mexican gov cyberattack stealing 100M+ records. Verizon’s 2025 DBIR showed third-party breaches doubled YoY.

Why this matters for Indian businesses

India’s DPDP Act 2023 holds data fiduciaries responsible for AI-caused data breaches. With 24% of Indian enterprises already deploying agentic AI (EY AIdea of India 2026) and TCS, Infosys, Wipro, Cognizant deploying 200,000+ Microsoft Copilot licences, understanding failure modes is critical before — not after — deployment.

Key deployment lesson

Deloitte Tech Trends 2026: only 11% of organisations have agents in production. APEX-Agents benchmark: even top models completed fewer than 25% of real-world tasks on first attempt. The common failure pattern: agents given write access to production systems without staging environments or human-in-the-loop approvals.

Submit an incident →