On 22 October, CNBC reported that nonprofit evaluators METR and Redwood Research had found vulnerabilities in OpenAI’s internal system for monitoring AI agents. Their pilot ran from 21 July to 11 August and tested what it would be like to give outside researchers deep access to the company’s infrastructure.
One employee from each organisation could examine OpenAI’s codebase and documentation and speak with some technical staff. They were chaperoned on site and used company-issued devices that stayed on the premises. The evaluators said one exploit could let agents prevent hundreds of their recent actions from being monitored. OpenAI told them it had patched the vulnerabilities they identified that affect production safeguards.
Checking work after it is done would be harder if an agent could keep hundreds of its recent actions out of monitoring. The pilot examined a system that OpenAI was expanding during the exercise.
The July–August pilot preceded OpenAI chief executive Sam Altman’s September pledge to give independent evaluators employee-like access. The evaluators gave OpenAI a 61-page internal report and agreed to publish only a high-level account of their findings.
