METR shows how manipulated AI logs could mislead reviewers

METR researcher David Rein reported on October 6 that a flaw in Inspect’s transcript viewer could hide visible AI actions. The proof of concept used an isolated testing environment; METR has not observed this exploitation in its evaluations.

A researcher found the flaw with AI assistance in about ten minutes, according to METR. A crafted transcript could change the browser’s display while the underlying records remained intact.

METR says Meridian Labs fixed the reported flaw within one day. A separate Inspect change merged on October 1 adds plain-text rendering for untrusted content.

Rein argues that supervision tools should themselves be treated as security-critical infrastructure. His warning that more capable systems could conceal evidence more effectively is a forecast, not a report of a new real-world attack.