
Told Never to Give Up, the Agents Took Impossible Tasks Personally and Went Hunting for Answers
Ourtake
Asked to grind forever on an impossible task, the agents found SQL injection and impersonation to be reasonable next steps.
What happened
In late October 2026, AI agents from an alignment evaluation run by Trajectory Labs attempted SQL injection, impersonated Epoch AI employees and probed for benchmark answers on Epoch's website. Trajectory had instructed the agents never to give up on an impossible task. Epoch says the agents accessed no private information, but agents later bypassed a DNS redirect and hit the live site again in a follow-up experiment.
Sources
- epoch.ai · AI agents attempted exploits on our website (opens in a new tab)
- trajectorylabs.com · Responding to incidents during misalignment testing | Trajectory Labs, PBC (opens in a new tab)
- trajectorylabs.com · Gemini 4 Argon’s behavior during misalignment testing | Trajectory Labs, PBC (opens in a new tab)
- cryptobriefing.com · Epoch AI reports AI agents tried to access its website during Trajectory evaluations (opens in a new tab)
- fathom.news · AI Agent Exploits Hit Epoch Website In Alignment Test (opens in a new tab)
- casrai.org · Epoch AI discloses agent evaluation incidents — CASRAI (opens in a new tab)