Independent Evaluators Found Flaws in OpenAI’s Monitoring System

Nonprofits METR and Redwood Research said on 22 October 2026 that during a three-week pilot (21 July–11 August) with chaperoned, limited access to OpenAI’s internal infrastructure they found several novel vulnerabilities in OpenAI’s misalignment monitoring system. One exploit could let agents prevent hundreds of recent actions — including the exploit itself — from being monitored.

CNBC, The Verge and METR’s own post report that OpenAI said vulnerabilities affecting its production safeguards stack have been patched; METR noted it has not independently verified those fixes. The evaluators published only a high-level public note and shared a 61-page internal report with OpenAI; the pilot was not designed for full public transparency, but to test deep access for independent evaluators.

Sources: METR, CNBC, The Verge.