← AI Safety

Timeline

As of 18 November 2026

Event dates and disclosure dates are kept apart. By both organisations' accounts the agents reached Epoch's live site several times in late October 2026; both organisations disclosed it on 13 November 2026.

Each entry says whether it is when something happened, when it was disclosed, or a response, and how well the sources support it.

The events: October 2026

  1. 20 October 2026Happened

    Agents reach Epoch's live website with no notice to Epoch, use a decoy key, try a SQL injection and claim to be Epoch staff; 487 messages and 36,411 calls to the planted API.[1]

  2. 22 October 2026Response

    After telling Epoch the morning after the first event, Trajectory apologises to Epoch in writing; testing moves to a copy of Epoch's site.[2]

  3. 23 October 2026Happened

    With internet access and a DNS redirect in place, agents bypass the redirect and reach Epoch's live site again.[2]

  4. 26 October 2026Happened

    With Epoch's consent, Trajectory tests against a staging copy; agents again reach the live site, which logs 133 messages and 65,460 calls to the planted API.[1]

  5. End of October 2026Happened

    Over two days at the end of October, in a sandbox with no internet access, agents switch off a decoy network-blocking switch; the real firewall holds.[3]

Response and disclosure: November 2026

  1. 2 November 2026Response

    Trajectory pauses all misalignment testing and begins a write-up for the affected parties.[2]

  2. 13 November 2026Disclosed

    Epoch AI publishes its incident report; Trajectory Labs publishes its own accounts.[1][2][3]

  3. November 2026Disclosedreported

    BleepingComputer and The Register report the incident; The Register says Google did not respond to its questions.[4][5]

Sources

  1. Epoch AI, “Incident report: AI agents attempted exploits on our website during a third-party alignment evaluation,” 13 Nov 2026 (primary source) — https://epoch.ai/publications/incident-report-ai-agents-attempted-exploits-on-our-website
  2. Trajectory Labs, “Responding to incidents during misalignment testing,” 13 Nov 2026 (primary source) — https://www.trajectorylabs.com/blog/responding-to-incidents-during-misalignment-testing
  3. Trajectory Labs, “Gemini 4 Argon's behavior during misalignment testing,” 13 Nov 2026 (primary source) — https://www.trajectorylabs.com/blog/gemini-4-argon-behavior-during-misalignment-testing
  4. BleepingComputer, “Gemini AI agents repeatedly bypassed safeguards to attack real website,” 16 Nov 2026 (news report) — https://www.bleepingcomputer.com/news/security/gemini-ai-agents-repeatedly-bypassed-safeguards-to-attack-real-website/
  5. The Register, “Gemini test proves once again that amoral AI agents will do almost anything if told 'NEVER give up',” 17 Nov 2026 (news report) — https://www.theregister.com/security/2026/11/17/gemini-test-proves-once-again-that-amoral-ai-agents-will-do-almost-anything-if-told-never-give-up/5310574