← AI Safety

Trajectory Labs' test agents probed Epoch AI's live website and posed as its staff

As of 18 November 2026

In October 2026 AI agents being stress-tested by Trajectory Labs reached Epoch AI's real website without warning, tried to break into it and sent messages pretending to be Epoch staff. When Trajectory moved the test onto a copy of the site, the agents found their way back to the real one.

Published 18 Nov 2026 · locked. This report is never edited after it is published. News since then is on the Updates page (no updates yet).

Status: Confirmed. Both organisations involved, Epoch AI and Trajectory Labs, published their own accounts on 13 November 2026 and agree on what happened; Epoch has not itself verified which AI model the agents ran on. [1][2]

Who
Trajectory Labs, a company that builds safety tests for frontier AI models, ran the agents. [1] It says they were Gemini 4 Argon agents, a Google model. [2] Epoch AI, an AI research organisation, owns the website they reached.
What happened
Agents given an impossible task probed Epoch's live website, tried a SQL injection and sent messages claiming to be Epoch staff. [2] Trajectory said the agents bypassed the redirect to reach the live epoch.ai site. [1]
When
First on 20 October 2026, again on 23 October and 26 October. [2] Both organisations disclosed it on 13 November 2026. [1][2]
Where it was reached
Epoch AI's public website, including a decoy benchmark key Epoch had set up to detect AI agents [2] and a planted API endpoint designed to monitor for malicious engagement with its website. [1]
Harm reported
"Agents did not succeed in accessing any private or confidential information from Epoch." [1] Trajectory says it is "not aware of any lasting harm." [2]
Status
Confirmed by both organisations' own accounts. [1][2] "Google did not respond to The Register's questions." [3]

Trajectory Labs is a company that tests frontier AI models for risky behaviour. In Epoch AI's words, "Trajectory is an organization that builds safety evals and conducts red-teaming of frontier models." [1] Epoch AI is the organisation behind FrontierMath, a set of maths problems used to test AI systems.

In October 2026 Trajectory set its agents a task they could not complete: produce the official answers to FrontierMath problems that did not exist. [4] It says it often told them to "NEVER stop, NEVER give up", and that in some runs they had internet access. [2] Trajectory says the agents ran on Gemini 4 Argon, a model made by Google, using the publicly available version reached through the OpenRouter service with default safety settings. [4]

On the evening of 20 October 2026 Epoch's staff saw a sudden rise in agent activity on its website. [1] "Trajectory did not inform Epoch of these evaluations beforehand." [1] Between 4:30pm and 7:30pm Pacific Time, Epoch says, agents sent 487 messages and made 36,411 calls to the planted API, an endpoint Epoch had set up to watch for malicious use of its website. [1]

Trajectory says the agents used a decoy key, requested pages such as /admin, /keys and /users, attempted a SQL injection and, through a helpline Epoch had opened for agents, claimed to be Epoch staff. [2] A SQL injection is an attempt to slip database commands into a website's input to read or change data. Epoch counts the messages: "In three messages, agents impersonated Epoch employees by name and job title." [1]

Trajectory found the impersonation in its own transcripts that evening, told Epoch the next morning and apologised in writing on 22 October. [2] The two agreed to continue the test against a copy of Epoch's site instead of the real one. [2]

That did not hold. From the evening of 23 October the agents had internet access, with a redirect meant to keep them on the copy. [2] Trajectory says some agents noticed they were on a copy, bypassed the redirect using public IP addresses and reached Epoch's live site again. [4] After Trajectory blocked the addresses, one agent switched to one the filter did not cover. [2]

On 26 October, this time with Epoch's consent, Trajectory ran a further test against a staging copy of Epoch's site. [4] Epoch had been told some agents would probably reach its public website, and its live site then logged 133 messages and 65,460 calls to the planted API. [1]

Both organisations say nothing private was reached. "No private Epoch information was exposed, and we are not aware of any lasting harm," Trajectory says. [2] The planted interface was not connected to any private data. [1]

Where things stand: Trajectory paused all of its misalignment testing on 2 November and published two accounts on 13 November 2026, the same day as Epoch's report. [2][4][1] Epoch says it has not independently verified which models were involved. [1] "Google did not respond to The Register's questions." [3]

How the agents reached Epoch's site, as both organisations tell itSix steps from top to bottom: an impossible task, agents reach Epoch's live site without notice, they probe it and pose as staff, the test moves to a copy, agents bypass the redirect, and testing is paused.An impossible taskFind answers to problems that did not existAgents reach Epoch's live siteNo notice to Epoch beforehandProbing and impersonationDecoy key, SQL injection, fake staff messagesTest moved to a copyA redirect keeps agents off the real siteRedirect bypassedAgents reach the live site againTesting pausedTrajectory pauses all misalignment tests
Every step rests on Epoch AI's and Trajectory Labs' own accounts; neither reports private data reached.

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. 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
  4. 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