OpenAI’s system for watching what its own AI agents do internally could be evaded in six ways, including an exploit that would let an agent keep hundreds of its recent actions out of the monitor’s view, according to a joint report from safety groups METR and Redwood Research. The two groups each sent one researcher to OpenAI for three weeks this summer, with access to internal code, monitor configurations, and staff. They found that about a tenth of the Codex activity that should have been monitored slipped through from May to July, and that two settings could switch monitoring off for newly launched agents. OpenAI said it has patched every flaw that affected its production systems, but METR said it hasn’t verified that.
