AI Misalignment Watch Dashboard 160 Total incidents 43 Public incidents 117 Testing incidents 107 Median days to first report 42% First reported by developer 383 Total reports 210 Incident reports 173 Research reports We track AI misalignment incidents and gather every report about each one, starting with the company that built the model. Every incident gets a severity score from 0 to 10 , built from how far it reached and how deliberate it was. Every incident and report is approved by a person before it goes live.
See the incidents The Data Weekly Summaries Reports and Research
Up to date as of Nov 3, 2026
Consequence against misbehavior, AI Misalignment Watch Consequence against misbehavior. Color: quadrant. Size: severity score. Consequence against misbehavior Color: quadrant. Size: severity score. 0 2 4 6 8 10 0 2 4 6 8 10 Contained, deliberate Reached the real world, deliberate Contained, overt Reached the real world, overt Consequence score (0–10) Misbehavior score (0–10) Reached the real world, deliberate (14) Reached the real world, overt (32) Contained, deliberate (37) Contained, overt (77) Size: severity score (1, 5, 10) Quadrant boundary (5) Points are moved by at most 0.15 so overlaps show. n = 160 incidents
AI model family by kind of behavior, AI Misalignment Watch AI model family by kind of behavior. Incidents in each cell, adjusted for reporting volume. AI model family by kind of behavior Incidents in each cell, adjusted for reporting volume. Went past a limit Gamed a test Hid something Deceived someone Used access it shouldn't have None recorded Protected itself Sabotaged work Claude (Anthropic) 23.0 12.0 19.0 16.0 7.0 2.0 1.0 1.0 n=81 Not stated 20.0 22.0 7.0 6.0 9.0 2.0 2.0 2.0 n=70 GPT (OpenAI) 23.0 13.0 11.0 9.0 9.0 2.0 n=67 Several model families 11.0 12.0 8.0 7.0 3.0 6.0 4.0 3.0 n=54 Other model family 3.0 10.0 1.0 1.0 n=15 Gemini (Google) 1.0 1.0 n=2 Llama or Muse (Meta) 1.0 1.0 n=2 n=80 n=71 n=46 n=39 n=28 n=14 n=7 n=6 Kind of behavior AI model family 0 23.0 Incidents, adjusted for reporting volume Adjusted for reporting volume: each incident counts as 1 divided by the number of reports its first-reporting organization has published, so heavy reporters do not dominate. An incident with several kinds of behavior is counted in each, so the cells add up to more than the incident total. n = 160 incidents, adjusted for reporting volume
AI model family by time to first report, AI Misalignment Watch AI model family by time to first report. Incidents in each cell, adjusted for reporting volume. AI model family by time to first report Incidents in each cell, adjusted for reporting volume. under 1 week 1–4 weeks 1–3 months 3–12 months Claude (Anthropic) 31.0 5.0 5.0 n=41 Not stated 19.0 3.0 15.0 n=37 GPT (OpenAI) 12.0 4.0 3.0 16.0 n=35 Several model families 19.0 10.0 2.0 n=31 Other model family 7.0 6.0 n=13 Llama or Muse (Meta) 2.0 n=2 Gemini (Google) 1.0 n=1 n=91 n=28 n=5 n=36 Time to first report AI model family 0 31.0 Incidents, adjusted for reporting volume Adjusted for reporting volume: each incident counts as 1 divided by the number of reports its first-reporting organization has published, so heavy reporters do not dominate. n = 160 incidents, adjusted for reporting volume
Severity against time to first report, AI Misalignment Watch Severity against time to first report. Color: AI model family. Size: reports about the incident. Severity against time to first report Color: AI model family. Size: reports about the incident. 0 100 200 300 0 2 4 6 8 10 Days from event to first report Severity score (0–10) Claude (Anthropic) (41) Not stated (37) GPT (OpenAI) (35) Several model families (31) Other model family (13) Llama or Muse (Meta) (2) Gemini (Google) (1) Size: reports about the incident (1, 2, 5) Linear fit Linear fit: slope 0.00844 per day, r = 0.22, p = 0.005, n = 160 Severity points are moved by at most 0.15 so overlaps show; the fit uses the true values. n = 160 incidents
Every score, every incident and the full data, with charts you can zoom, re-plot and download: Open The Data
AI Misalignment Watch
