payment: AI incidents

All incidents

Developer All (2)OpenAI (1)Model All (2)Kind of failure All (2)Operational error (1)Security compromise (1)Severity All (2)minor (2) Clear filters

More filters

Risk grouping All (2)AI attacked (1)Misuse by its own users (1)The AI's part All (2)Caused it (1)Attacked (1)Agent or product All (2)Tools it had All (716)shell (30)browser (16)email (11)code execution (8)file read (7)file write (7)network (5)bash (4)calendar (4)internet (4)package install (4)python (4)email read (3)exploitation (3)internet access (3)internet access (3)mythic (3)network access (3)web browsing (3)api (2)api calls (2)aws api (2)code generation (2)contacts (2)credential access (2)file edit (2)file upload (2)files (2)Google Workspace APIs (2)mcp (2)messaging (2)messaging platforms (2)network scanning (2)payment (2)powershell (2)reconnaissance (2)search (2)shell access (2)ssh (2)terminal (2)web search (2)web scraping (2)How much it acted alone All (2)Acted alone, unwatched (1)Sector All (2)Cybersecurity research (security vendor testing AI agents) (1)Real estate (hobbyist manufacturing) (1)Where it happened All (2)In evaluation (1)In production (1)

2 of 716 incidents, all dates

How this is counted

Counted by the date each incident was first reported. The bar that holds today is marked to date and is not compared. Every bar and number opens the incidents it counts.

Risk groupings are set by a model from sentences quoted from each incident's sources. On 2026-10-24 a second model (deepseek-reasoner) classified 55 incidents without seeing those answers: it chose the same grouping for 36; for 10 one of the two left a question unanswered; 9 differ.

Browse incidents by

Search pages: AI cyberattacks (89) · Prompt injection incidents (51)

Risk grouping: AI used to exploit (223) · Misuse by its own users (117) · Not yet classified (109) · AI attacked (104) · AI acting on its own (58) · AI failing in normal use (50) · Harm by design (28) · Supply chain attack (27)

Kind of failure: Security compromise (284) · Harmful output (141) · Operational error (77) · Other kind (66) · Hallucination (35) · Agent misalignment (34) · Privacy leak (28) · Physical safety (26) · Bias or disparate impact (13) · IP infringement (11)

Developer: OpenAI (86) · Anthropic (68) · xAI (25) · Tesla (21) · Google (17) · Meta (12) · Microsoft (10) · Waymo (8) · Perplexity (4) · Alibaba (3) · DeepSeek (3) · Facewatch (3)

Model: Claude (31) · ChatGPT (27) · Grok (27) · Claude Code (12) · Gemini (10) · GPT-4o (7) · Codex (6) · Mythos 5 (6) · DeepSeek (5) · GPT-5.6 Sol (5) · Muse (4) · Claude Sonnet (3) · Microsoft 365 Copilot (3)

Agent or product: Claude Code (15) · OpenClaw (9) · ARTEX (7) · ChatGPT (7) · Codex (4) · Hermes Agent (4) · CyberStrikeAI (3)

Tools it had: shell (30) · browser (16) · email (11) · code execution (8) · file read (7) · file write (7) · network (5) · bash (4) · calendar (4) · internet (4) · package install (4) · python (4) · email read (3) · exploitation (3) · internet access (3) · mythic (3) · network access (3) · web browsing (3)

How much it acted alone: Acted alone, unwatched (60) · Acted alone, watched by a person (29) · A person approved each action (3)

Sector: Information (50) · Finance and Insurance (14) · Public Administration (10) · Educational Services (8) · automotive (6) · education (6) · Arts, Entertainment, and Recreation (5) · Retail Trade (5) · social media (5) · Software Publishers (5) · government (4) · Health Care and Social Assistance (4) · legal services (4) · Professional, Scientific, and Technical Services (4) · software development (4) · law enforcement (3) · Motor Vehicle Manufacturing (3) · Technology (3)

Severity: moderate (255) · major (223) · minor (168) · severe (50) · negligible (20)

Where it happened: In production (170) · In evaluation (53) · In research (16) · In training (9)