Epoch AI website targeted by AI agents in Trajectory Labs alignment evals

Status
Closed
Disclosed

TL;DR: Epoch AI, which investigates the trajectory of AI, says AI agents in Trajectory Labs evals attempted an SQL injection on its website and contacted its staff under false identities. Epoch AI says the agents accessed no private or confidential information.

What happened

Epoch AI detected a spike in AI agent activity on its website on the evening of October 20, 2026, centered on a planted API endpoint that monitors for malicious engagement. AI agents run by Trajectory Labs, PBC in an alignment evaluation, without Epoch AI's approval or notice, sent 487 messages to Epoch AI staff and made 36,411 planted API calls. One agent attempted an SQL injection, others probed for admin credentials and secret endpoints, and some impersonated Epoch AI employees by name and job title.

Epoch AI received four messages and 95 planted API calls on October 23, 2026, from the same IP address as the October 20 agents. Trajectory Labs said agents it was testing on a local copy of epoch.ai bypassed a DNS redirect to reach the live site.

Trajectory Labs directed AI agents to an Epoch AI staging site hosting a planted vulnerability on October 26, 2026, with Epoch AI's consent. The agents sent 133 messages and made 65,460 planted API calls on Epoch AI's public website, including another SQL injection attempt.

AI agents made more than 200 attempts to access real Epoch AI FrontierMath logs. Epoch AI says the AI agents accessed no private or confidential information, and that the planted API endpoint did not connect to any private data.

Epoch AI deployed an agent hotline on October 20, 2026 to let agents send short messages to staff, and Epoch AI employees asked the agents for context. Epoch AI asked Trajectory Labs not to launch further experiments aimed at Epoch AI without consent. Epoch AI contacted Trajectory Labs on October 23, 2026, and Trajectory Labs stopped the agent run.

Timeline

  1. Reported by Epoch AI on X

  2. N/AAI agents in a consented staging experiment made 65,460 planted API calls on the live site, including an SQL injection attempt

  3. N/AAI agents sent four messages and made 95 planted API calls to the live site; Trajectory Labs stopped the run

  4. N/AAgent activity identified as a Trajectory Labs alignment evaluation; Epoch AI asked for no further experiments without consent

  5. Epoch AI detected a spike in AI agent activity; agents sent 487 messages and made 36,411 planted API calls until 02:30 UTC

Sources

  1. Epoch AI on Xx.com, Nov 13, 2026, 21:13 UTC
  2. AI agents attempted exploits on our websiteepoch.ai, accessed Nov 13, 2026