Evaluators say 'vulnerabilities' found with OpenAI monitoring system

Oct. 22 (UPI) -- Nonprofit evaluators said Thursday they have found "vulnerabilities" with an OpenAI monitoring system in a trial run of how the company will work with third-party safety organizations going forward. According to METR and Redwood Research, their pilot review of OpenAI's system was a test to see what it would look like for the company to provide "a deep level of access" to its internal infrastructure. The organizations said one employee from METR and one from Redwood had access to OpenAI's codebase, documentation and some technical staff between July 21 and Aug. 11. The METR and Redwood staffers were chaperoned while using OpenAI devices which "never left the premises." Several "novel vulnerabilities," including one that would allow agents to prevent "hundreds" of recent actions from being monitored…