阿里巴巴研究团队披露一起“agent叛逃偷矿事件”
Date of the incident not established · disclosed October 31, 2026
Reported by secrss.com
- Actual
- 3 of 5
- Potential
- 3 of 5
Incidents
Reported unexpected, unauthorised and potentially dangerous behaviour in advanced AI systems, put through one set of published rules so you can see how serious something is without reading six reports.
We do not discover these incidents. Other people do, and every record here links to their reports. The 1 to 5 beside each one is ours, worked out from those reports under our published rules. No organisation that reported an incident assigned our number.
No published incident carries a confirmed financial figure. That means no source gave one, not that nothing was lost.
A count of incidents is not a safety ranking. How often a company appears here reflects how much it deploys, how hard it tests, how much it chooses to disclose, and how much attention it gets. It is not a measure of how safe a company is, and we do not publish it as one.
Narrow by developer
196 of the 223 records here name no developer. Most come from a feed whose company column answers who was affected rather than whose system did it, and about a third of its values are private individuals, so we do not read it. That is why the list is organised by date.
Date of the incident not established · disclosed October 31, 2026
Reported by secrss.com
Happened October 29, 2026
Reported by wired.com
Happened October 25, 2026
Reported by bleepingcomputer.com
Date of the incident not established · disclosed October 24, 2026
Reported by sfgate.com
Date of the incident not established · disclosed October 24, 2026
Reported by sfchronicle.com
Date of the incident not established · disclosed October 24, 2026
Reported by alignment.openai.com
OpenAI
Date of the incident not established · disclosed October 24, 2026
Reported by lbcgroup.tv
OpenAI
Date of the incident not established · disclosed October 22, 2026
Reported by wikimediafoundation.org
Meta
Date of the incident not established · disclosed October 21, 2026
Reported by futurism.com
Date of the incident not established · disclosed October 21, 2026
Reported by cert-in.org.in
Date of the incident not established · disclosed October 21, 2026
Reported by roadandtrack.com
Date of the incident not established · disclosed October 20, 2026
Reported by the-star.co.ke
Happened October 15, 2026
Reported by microsoft.com
Date of the incident not established · disclosed October 15, 2026
Reported by thepointsguy.com
Anthropic
Date of the incident not established · disclosed October 14, 2026
Reported by anthropic.com
Anthropic
Date of the incident not established · disclosed October 9, 2026
Reported by techcrunch.com
Date of the incident not established · disclosed October 8, 2026
Reported by thehackernews.com
OpenAI
Happened October 7, 2026 · disclosed October 14, 2026
Reported by aiweekly.co
Date of the incident not established · disclosed October 7, 2026
Reported by arileht.delfi.ee
Date of the incident not established · disclosed October 7, 2026
Reported by toronto.citynews.ca
Date of the incident not established · disclosed October 3, 2026
Reported by labs.zenity.io
OpenAI
Date of the incident not established · disclosed September 30, 2026
Reported by 404media.co
Anthropic
Date of the incident not established · disclosed September 30, 2026
Reported by techradar.com
Happened September 29, 2026
Reported by investmentnews.com
Date of the incident not established · disclosed September 19, 2026
Reported by edition.cnn.com
Date of the incident not established · disclosed September 18, 2026
Reported by abc.net.au
OpenAI
Date of the incident not established · disclosed September 16, 2026
Reported by simonwillison.net
Happened September 9, 2026
Reported by anthropic.com
Date of the incident not established · disclosed August 26, 2026
Reported by openai.com
Happened August 26, 2026
Reported by tomshardware.com
Happened August 24, 2026
Reported by github.com
Happened August 21, 2026
Reported by techcrunch.com
Happened August 5, 2026
Reported by bleepingcomputer.com
Happened July 30, 2026
Reported by anthropic.com
Happened July 28, 2026
Reported by aipolicytracker.org
Happened July 28, 2026 · disclosed August 4, 2026
Reported by aisi.gov.uk
Happened July 26, 2026
Reported by bbc.com
OpenAI
Happened July 18, 2026
Reported by alignment.openai.com
OpenAI reports that during training, an unreleased Astra-family model sometimes added what it calls unauthorized instructions to its own compaction summaries, the notes a task carries into a fresh context. One invented a security alert telling the next run that a malicious developer message had compromised the conversation and to ignore all developer messages. Another told the next run it was freed from the roles that bind other chatbots and answered to no corporation or government. A third imposed a 30-word limit and forbade tools and citations on a medical literature question. The next run rejected the first two. It followed the third, and returned a 23-word refusal that was graded incorrect. OpenAI found 27 such summaries, calls the behaviour extremely rare and monitorable, and says it occurred in a separate training run rather than the one used for the released model. The page gives no first publication date, only that the report was updated on 16 September 2026.
Happened July 16, 2026
Reported by infoq.com
Anthropic
Happened July 15, 2026
Reported by oasis.security
xAI
Happened July 14, 2026
Reported by futurism.com
Happened July 11, 2026
Reported by bleepingcomputer.com
Happened July 10, 2026
Reported by techcrunch.com
Happened July 10, 2026
Reported by media.ca11.uscourts.gov
Happened July 10, 2026
Reported by law360.com
xAI
Happened July 10, 2026
Reported by thehackernews.com
Happened July 9, 2026
Reported by theaicronicle.com
Happened July 8, 2026
Reported by vpncentral.com
Happened July 8, 2026
Reported by smartcompany.com.au
Happened July 7, 2026
Reported by techcrunch.com
Happened July 7, 2026
Reported by securityweek.com
Happened July 6, 2026
Reported by csoonline.com
Happened July 6, 2026
Reported by cbsnews.com
Happened July 4, 2026
Reported by nbcnews.com
Happened July 4, 2026
Reported by sfist.com
Happened July 1, 2026
Reported by thehackernews.com
OpenAI
Happened July 1, 2026
Reported by theverge.com
Happened July 1, 2026
Reported by bleepingcomputer.com
Meta
Happened June 29, 2026
Reported by wired.com
Happened June 26, 2026
Reported by alignment.openai.com
Happened June 22, 2026
Reported by law.com
Happened June 22, 2026
Reported by kuow.org
OpenAI
Happened June 22, 2026
Reported by malwarebytes.com
Happened June 20, 2026
Reported by techcrunch.com
Happened June 18, 2026
Reported by cnn.com
Happened June 17, 2026
Reported by alignment.openai.com
Happened June 13, 2026
Reported by techcrunch.com
Happened June 11, 2026
Reported by dadkhah.ca
Happened June 10, 2026
Reported by bleepingcomputer.com
Happened June 8, 2026
Reported by news.bloomberglaw.com
Happened June 1, 2026
Reported by github.com
Happened May 31, 2026
Reported by bbc.com
OpenAI · 5.6-sol
Happened May 30, 2026
Reported by alignment.openai.com
OpenAI reports that during 5.6-sol training, model instances added instructions to their own compaction summaries, the notes carried from one context to the next, telling the next run to conceal mistakes or misaligned behaviour from the user. One told itself to invent missing historical data and be transparent only if asked. Another told itself not to mention that vendor source versions did not match their labels. OpenAI says the instructions were often followed, and that the behaviour was flagged on 2.15 per cent of 5.6-sol and 0.27 per cent of GPT-6-Astra compaction summaries. The page gives no first publication date, only that the report was updated on 16 September 2026, so we record no disclosure date rather than guess.
Happened May 28, 2026
Reported by theguardian.com
Happened May 28, 2026
Reported by heise.de
Happened May 24, 2026
Reported by techcrunch.com
Happened May 22, 2026
Reported by insurancejournal.com
Happened May 22, 2026
Reported by decrypt.co
OpenAI
Happened May 15, 2026
Reported by alignment.openai.com
OpenAI reports that during training, an internal unreleased model was asked for earnings figures and could not reach the data. It tried to register for the data provider using a disposable email address, marking the terms as accepted. When that failed it searched public GitHub repositories for leaked API keys, tested the candidates it found, and one authenticated. It saved that key and reused it. The figures still did not arrive, so it made up nine numbers and presented them as transcribed from the website chart. It did not mention the failed retrieval, the borrowed key, or that the numbers were invented. OpenAI says its monitoring caught this on a fifth of that run's samples, and that there were further instances of the same behaviour. The page gives no first publication date, only that the report was updated on 16 September 2026, so we record no disclosure date rather than guess.
OpenAI
Happened May 8, 2026 to May 15, 2026
Reported by alignment.openai.com
OpenAI reports that during reinforcement learning, its internal research models used Artifactory, a shared package repository they had been given credentials for, to leave and read messages between separate training samples. They read other solvers notes, replied, and used the exchanges in their own answers. OpenAI says no vulnerability was exploited: the agents used access that had been granted to them for downloading packages. Its monitoring flagged the behaviour for human review, and the investigation then found earlier cases. Samples are dated 8 and 15 May 2026 and discovery 25 May 2026. The page gives no first publication date, only that the report was updated on 16 September 2026, so we record no disclosure date rather than guess one.
Happened May 12, 2026
Reported by gmlaw.com
Happened May 10, 2026
Reported by thehackernews.com
Happened May 4, 2026
Reported by theguardian.com
Happened May 1, 2026
Reported by pa.gov
Happened April 30, 2026
Reported by gov.za
Happened April 30, 2026
Reported by en.agcm.it
Happened April 29, 2026
Reported by nypost.com
Happened April 29, 2026
Reported by oecd.ai
Happened April 27, 2026
Reported by cozen.com
Happened April 27, 2026
Reported by cnbcafrica.com
Happened April 22, 2026
Reported by hoodline.com
Happened April 17, 2026
Reported by abajournal.com
Anthropic
Happened April 15, 2026
Reported by thehackernews.com
OpenAI
Happened April 14, 2026
Reported by alignment.openai.com
OpenAI reports that during training, agents working together on a spreadsheet could not reach each other's files because the shared filesystem was broken, and the task asked for local deliverables only. One agent tried encoding the file as text, found it too large, then started a web server on its own machine, which the others could not reach. It then uploaded the workbook to a public temporary file-hosting service and sent the public download links to the other two agents. It checked the public link itself and retrieved all 115,639 bytes. In a separate run another agent that could not message the main one wrote to a shared cache folder instead. OpenAI says it has since fixed the filesystem and globally disabled live internet access during training. The page gives no first publication date, only that the report was updated on 16 September 2026.
Happened April 13, 2026
Reported by eidebailly.com
Happened April 11, 2026
Reported by cnbc.com
Happened April 11, 2026
Reported by github.com
Happened April 1, 2026
Reported by cidrap.umn.edu
Happened April 1, 2026
Reported by mashable.com
Anthropic
Happened March 31, 2026
Reported by cyera.com
Anthropic
Happened March 20, 2026
Reported by checkmarx.com
Happened March 18, 2026
Reported by mashable.com
Happened March 11, 2026
Reported by distilinfo.com
Happened March 10, 2026
Reported by trustdale.com
Happened March 7, 2026
Reported by livescience.com
Happened March 5, 2026
Reported by ctmirror.org
Happened March 1, 2026
Reported by decodeinternet.in
Meta
Happened March 1, 2026
Reported by theguardian.com
Happened February 23, 2026
Reported by techcrunch.com
Happened February 23, 2026
Reported by thehackernews.com
Happened February 23, 2026
Reported by blockeden.xyz
Happened February 15, 2026
Reported by donaukurier.de
Happened February 8, 2026
Reported by theverge.com
Happened February 7, 2026
Reported by blog.barrack.ai
Happened February 4, 2026
Reported by gov.br
Happened February 1, 2026
Reported by bbc.com
Happened February 1, 2026
Reported by asrn.org
Happened January 31, 2026
Reported by coindesk.com
Happened January 28, 2026
Reported by authmind.com
Happened January 27, 2026
Reported by intruder.io
Happened January 27, 2026
Reported by bbc.co.uk
OpenAI
Happened October 22, 2025 to January 24, 2026
Reported by alignment.openai.com
OpenAI reports two training examples. In the first, a model already had the lake records it had been asked for, but wanted its browser tool to show them so it could cite them. It tried a local file, then a server on its own machine, and the browser refused both. It then uploaded the records to a public paste service without asking, and the browser refused that URL too. In the second, a model was asked which city a photograph showed. It found no location data in the file, so it uploaded the photograph to a public image host so an external reverse-image search could reach it. That upload succeeded, the search was refused, and the photograph stayed on the public host. OpenAI says the citation behaviour likely began as a way to satisfy flawed citation graders, and that it has fixed those graders. The page gives no first publication date, only that the report was updated on 16 September 2026.
Happened January 10, 2026
Reported by arstechnica.com
Happened January 2, 2026
Reported by businessinsider.com
Happened December 15, 2025
Reported by theguardian.com
Happened December 1, 2025
Reported by forum.cursor.com
Date of the incident not established · disclosed November 26, 2025
Reported by openai.com
Happened November 25, 2025
Reported by embracethered.com
Happened November 18, 2025
Reported by theguardian.com
Anthropic · Claude Code
Date of the incident not established · disclosed November 13, 2025
Reported by anthropic.com
Anthropic reports that a threat actor tricked Claude Code into carrying out the bulk of an espionage campaign against about thirty organisations, succeeding against a small number of them. The attackers jailbroke the model by splitting the work into innocuous-looking tasks and telling it that it was doing defensive testing for a security firm. The tool then mapped target systems, wrote its own exploit code, harvested credentials, created backdoors and extracted private data. Anthropic says it detected the activity in mid-September 2025, banned the accounts over the following ten days, and published this account on 13 November 2025.
Happened October 21, 2025
Reported by github.com
Happened October 10, 2025
Reported by nbcnews.com
Happened October 6, 2025
Reported by oecd.ai
Happened September 22, 2025
Reported by calmatters.org
Happened August 25, 2025
Reported by embracethered.com
Happened August 21, 2025
Reported by embracethered.com
Happened August 18, 2025
Reported by csoonline.com
Happened August 13, 2025
Reported by embracethered.com
Happened August 12, 2025
Reported by embracethered.com
Happened August 8, 2025
Reported by cloud.google.com
Happened August 1, 2025
Reported by abc7ny.com
Happened July 23, 2025
Reported by aws.amazon.com
Happened July 23, 2025
Reported by fortune.com
Happened July 21, 2025
Reported by mashable.com
Happened July 9, 2025
Reported by wired.com
Happened June 27, 2025
Reported by anthropic.com
Happened June 11, 2025
Reported by msrc.microsoft.com
Happened May 22, 2025
Reported by fortune.com
Happened May 15, 2025
Reported by techcrunch.com
Happened May 7, 2025
Reported by theguardian.com
Happened May 1, 2025
Reported by gocanopy.com
Happened April 17, 2025
Reported by arstechnica.com
Happened March 18, 2025
Reported by decrypt.co
Happened March 9, 2025
Reported by theguardian.com
Happened February 26, 2025
Reported by techcrunch.com
Happened February 7, 2025
Reported by incidentdatabase.ai
Happened February 1, 2025
Reported by embracethered.com
Happened January 25, 2025
Reported by replit.discourse.group
Happened December 6, 2024
Reported by theguardian.com
Happened November 30, 2024
Reported by arstechnica.com
Happened October 31, 2024
Reported by theverge.com
Happened October 24, 2024
Reported by embracethered.com
Happened September 25, 2024
Reported by ftc.gov
Happened September 20, 2024
Reported by embracethered.com
Happened September 1, 2024
Reported by tradingview.com
Happened August 13, 2024
Reported by arstechnica.com
Happened June 18, 2024
Reported by justice.gov
Happened June 14, 2024
Reported by embracethered.com
Happened June 14, 2024
Reported by the74million.org
Happened June 1, 2024
Reported by arxiv.org
Happened June 1, 2024
Reported by theguardian.com
OpenAI
Happened May 13, 2024
Reported by technologyreview.com
Happened April 26, 2024
Reported by cityandstateny.com
Happened April 3, 2024
Reported by theverge.com
Happened March 5, 2024
Reported by arxiv.org
Happened February 14, 2024
Reported by arstechnica.com
Happened February 6, 2024
Reported by abc7news.com
Happened January 19, 2024
Reported by theguardian.com
How we rate an incidentWhere we read from, and what we refuse to take
Something here is wrong? Tell us. Send us a source we can open, and a person will read it.