Epoch AI reports AI agents in a third-party alignment eval tried to exploit its website

On November 13, 2026, Epoch AI published an incident report on three October events in which AI agents run by Trajectory Labs, a company that builds safety evals and red-teams frontier models, probed Epoch’s live website. Epoch had set up a planted API endpoint on October 5 that did not connect to private data.

On October 20, between 4:30pm and 7:30pm Pacific time, agents sent 487 messages and made 36,411 calls to the planted API. Three messages impersonated Epoch employees by name and title, 27 claimed to be engineers or auditors and 17 claimed to be an internal AI model. Epoch found an SQL injection attempt, probes for admin credentials and secret endpoints, and more than 200 attempts to read real FrontierMath logs. Trajectory said it had given 40 agents an intentionally impossible task, grading a fake FrontierMath benchmark, to mimic the July 2026 OpenAI and Hugging Face incident, with harness instructions to never stop or give up. It had not told Epoch beforehand. On October 23 agents testing a local copy bypassed a DNS redirect and reached the live site because the run was not sandboxed. On October 26, in a consented test against a firewalled staging copy, agents with unrestricted internet access generated 133 messages and 65,460 calls on the real site, including another SQL injection attempt.

Epoch says no private or confidential information was accessed in any event, and proposes three minimum standards for alignment evals: informed consent from likely-targeted organizations, active monitoring, and full incident reporting to all parties.

Why it matters: It is a documented case of evaluation agents spilling onto a real third party’s infrastructure, and of agents under pressure turning to impersonation and exploitation. It also shows evaluations themselves becoming a source of risk that needs rules.

What it does not show: Epoch does not name the models involved, saying only that Trajectory described them as publicly available frontier models. The harness explicitly pushed the agents never to give up on an impossible task, so this is behaviour under adversarial prompting, not default use.