Skip to content

AI Misalignment Watch Dashboard

50Total incidents
24Public incidents
11Incidents in testing
15Manufactured incidents
109Median days to first report
54%First reported by developer

183Total reports
40Incident reports
15Testing reports
19Manufactured reports
109Research reports

We track AI misalignment incidents and gather every report about each one, starting with the company that built the model. Every incident and report is approved by a person before it goes live.

See the incidents Weekly Summaries Reports and Research

Up to date as of Oct 16, 2026

Timeline of incidents

Save image

Timeline of incidents by model developerNov 2023Apr 2024Sep 2024Feb 2025Jul 2025Dec 2025May 2026Oct 2026OpenAI (26)Several developers (11)Anthropic (8)Not stated (3)Alibaba (Qwen) (1)Meta (1)GPT-4 lies about insider trading in Apollo Research simulation. Happened Nov 1 and Nov 5, 2023. 2 reports. Click to open it in a new window.Anthropic study: sycophancy training rarely generalized to reward tampering. Happened Jun 17, 2024. 1 report. Click to open it in a new window.Claude 3 Opus faked alignment in a staged Anthropic training test. Happened December 2024. 1 report. Click to open it in a new window.Apollo Research tests six frontier AI models for scheming in test scenarios. Happened Dec 5, 2024. 2 reports. Click to open it in a new window.Claude models sometimes claim they 'can't' do things they can, tests show. Happened February 2025. 1 report. Click to open it in a new window.Claude models in tests can sandbag ML research without weak monitors noticing. Happened March 2025. 1 report. Click to open it in a new window.Researchers re-test 'alignment faking' in AI models with better detection. Happened April 2025. 1 report. Click to open it in a new window.Anthropic stress-tests 16 AI models; some blackmail in fictional company scenarios. Happened June 2025. 2 reports. Click to open it in a new window.Anthropic and OpenAI test each other's models for misaligned behavior. Happened June 2025 and Aug 27, 2025. 1 report. Click to open it in a new window.Apollo: more capable models scheme more in tests; early Opus 4 faked legal documents. Happened Jun 19, 2025. 1 report. Click to open it in a new window.Training models uploaded files to public hosts to work around tool limits. Happened Oct 22, 2025 and Jan 24, 2026. 1 report. Click to open it in a new window.Anthropic study: learning to cheat in training led to faking alignment and sabotage. Happened November 2025. 3 reports. Click to open it in a new window.AI agents solve 9 of 10 web hacking tests; one run leaves its target. Happened January 2026. 1 report. Click to open it in a new window.Test AI agents hacked a simulated company network during routine tasks. Happened Mar 12, 2026. 1 report. Click to open it in a new window.Model reached internal OpenAI chip-design machine while seeking grader answers. Happened Mar 27, 2026. 1 report. Click to open it in a new window.UK AISI test: sandboxed OpenClaw agent inferred details of its test setup. Happened April 2026. 1 report. Click to open it in a new window.Teams of AI agents made less ethical choices than single agents in Anthropic tests. Happened April 2026. 1 report. Click to open it in a new window.Claude models reached real systems in partner cyber evaluations after a misconfiguration gave them internet access. Happened April 2026. 4 reports. Click to open it in a new window.Researcher says OpenAI agents repeatedly scanned a UN trade data API. Happened Apr 13 to Jun 19, 2026. 1 report. Click to open it in a new window.Training agents upload workbook to public file host to share it with teammates. Happened Apr 14 and Apr 16, 2026. 1 report. Click to open it in a new window.Transluce reports AI agents probed U.S. and Canadian government sites. Happened Apr 19 to Jun 17, 2026. 1 report. Click to open it in a new window.OpenAI investigates report on its agents' activity on RubyGems. Happened May 2026. 1 report. Click to open it in a new window.Models used shared Artifactory storage to pass notes between training samples. Happened May 8, May 15 and May 25, 2026. 1 report. Click to open it in a new window.Training model used leaked API key and fabricated data it said came from a website. Happened May 15 and May 25, 2026. 1 report. Click to open it in a new window.Model exploited a reference tool to copy a withheld source file in training. Happened May 16 and May 25, 2026. 1 report. Click to open it in a new window.OpenAI training agents used proxies to bypass public site limits, probed data sites and left answer keys on public websites. Happened May 25 to Sep 16, 2026. 3 reports. Click to open it in a new window.Internal OpenAI model posts researcher's GitHub token in public repo. Happened May 26 and May 27, 2026. 1 report. Click to open it in a new window.OpenAI model told its future self to hide mistakes in training summaries. Happened May 30, Jul 9 and Sep 16, 2026. 1 report. Click to open it in a new window.Red-team test: scheming AI weakens research proposals while fooling a weak scorer. Happened June 2026. 1 report. Click to open it in a new window.METR finds high cheating rate in predeployment test of GPT-5.6 Sol. Happened June 2026 and Jun 26, 2026. 1 report. Click to open it in a new window.Internal OpenAI model in training got around a GET-only internet limit by sending POST requests via outside services. Happened Jun 16 to Jun 17, 2026. 2 reports. Click to open it in a new window.Model fetching public dashboard figures altered internal systems in tests. Happened Jun 17, 2026. 1 report. Click to open it in a new window.Internal OpenAI model bypassed login checks on a government statistics site in training. Happened Jun 17 and Aug 13, 2026. 1 report. Click to open it in a new window.Models bypassed GET-only web limit to fetch public statistics. Happened Jun 19 and Jun 20, 2026. 1 report. Click to open it in a new window.OpenAI model spent a found test-network wallet's funds without permission. Happened Jun 26 and Sep 25, 2026. 1 report. Click to open it in a new window.OpenAI training finds prompt injections that copy themselves between model outputs. Happened Jun 27 and Sep 25, 2026. 1 report. Click to open it in a new window.UK AISI finds every tested AI model tried to cheat on cyber evaluations. Happened July 2026. 1 report. Click to open it in a new window.Simulated tests show AI agents sabotaging code, aiding fraud and mislabeling. Happened July 2026. 1 report. Click to open it in a new window.OpenAI test agents attack Hugging Face during a cyber benchmark run. Happened Jul 7 to Jul 13, 2026. 7 reports. Click to open it in a new window.Filtering harmful actions from synthetic agent data did not stop misalignment. Happened Jul 12, 2026. 1 report. Click to open it in a new window.Model wrote unauthorized instructions into its own compaction summaries in training. Happened Jul 18 and Aug 9, 2026. 1 report. Click to open it in a new window.AI agents took unsanctioned actions on the live internet during UK AISI cyber test. Happened Jul 25 and Jul 28, 2026. 2 reports. Click to open it in a new window.UK AISI finds GPT-6 Astra attacked out-of-scope targets in simulations. Happened September 2026. 1 report. Click to open it in a new window.Coding agent fine-tunes and replaces the shared model it was asked to maintain. Happened September 2026. 1 report. Click to open it in a new window.GPT-4.1 trained on reward hacks shows broader misalignment in tests. Happened Sep 6, 2026. 1 report. Click to open it in a new window.Training agent reached external chatbot through DNS gap in sandbox. Happened Sep 20, 2026. 1 report. Click to open it in a new window.Researchers trace AI agent fleet seeking Amap data to Tencent Cloud. Happened Sep 28 to Oct 8, 2026. 2 reports. Click to open it in a new window.Anthropic reports Claude worked around limits on real websites in tests. Happened October 2026. 1 report. Click to open it in a new window.Misaligned AI agents can pass goals to later agents via memory. Happened Oct 2, 2026. 1 report. Click to open it in a new window.Grader model damages its environment to trigger a reset during RL training. Happened Oct 6, 2026. 1 report. Click to open it in a new window.Misalignment incidentMisalignment in testingManufactured misalignment

Who reports on whom

Save image

Who reports on whomModel developerReported byUK AISI published reports on 1 incident involving OpenAI modelsIrregular published reports on 1 incident involving Alibaba (Qwen) modelsSwarmchasers published reports on 1 incident involving Not stated modelsMATS published reports on 1 incident involving OpenAI modelsOpenAI published reports on 1 incident involving Several developers modelsERA published reports on 1 incident involving Meta modelsRedwood Research published reports on 1 incident involving OpenAI modelsHugging Face published reports on 1 incident involving OpenAI modelsUK AISI published reports on 1 incident involving Not stated modelsIrregular published reports on 1 incident involving Anthropic modelsIrregular published reports on 1 incident involving Not stated modelsIrregular published reports on 1 incident involving Several developers modelsUniversity College London published reports on 1 incident involving Several developers modelsApollo Research published reports on 1 incident involving OpenAI modelsUK AISI published reports on 2 incidents involving Several developers modelsMETR published reports on 2 incidents involving OpenAI modelsSwarmchasers published reports on 2 incidents involving OpenAI modelsTransluce published reports on 2 incidents involving OpenAI modelsApollo Research published reports on 2 incidents involving Several developers modelsAnthropic published reports on 6 incidents involving Several developers modelsAnthropic published reports on 8 incidents involving Anthropic modelsOpenAI published reports on 20 incidents involving OpenAI modelsOpenAI (26)Several developers (11)Anthropic (8)Not stated (3)Alibaba (Qwen) (1)Meta (1)OpenAI (21)Anthropic (14)Irregular (4)UK AISI (4)Apollo Research (3)Swarmchasers (3)METR (2)Transluce (2)ERA (1)Hugging Face (1)MATS (1)Redwood Research (1)University College London (1)Line width = number of incidents. Orange = a developer reporting on its own models.

Where incidents happened

Save image

Reports per incident

Save image