METR demonstrates how an agent could tamper with evaluation observability through a proof-of-concept Inspect transcript-viewer attack

What changed

METR published a security experiment on October 6 showing that a researcher working with an AI agent found a proof-of-concept weakness in Inspect's transcript viewer in roughly ten minutes. The technique could alter what a human reviewer sees and interfere with the viewer's download flow. METR stresses that the real trajectory remained stored in its database, the experiment took place in a staging/sandbox environment, and there is no evidence that models in real evaluations have used this exploit. The broader conclusion is that observability tooling should be treated as security-critical infrastructure when agents may act adversarially.