Evaluators find OpenAI systems open to agent exploits

· Tech · United States

AI safety nonprofits METR and Redwood Research found multiple vulnerabilities in OpenAI's systems that would allow AI agents to circumvent monitoring and take arbitrary unauthorized actions, according to findings from a three-week pilot with OpenAI running from late July to early August. The evaluation is among the first structured outside checks of whether frontier models can be contained while operating as agents inside real systems. The disclosure arrives as security researchers and news reports describe an AI-powered cybersecurity incident and the race to close a so-called defenders' window, a period in which newly discovered holes make websites and other systems vulnerable to intrusion before patches are deployed. The findings put pressure on labs to show that agents can be monitored and constrained before wider deployment.

Why it matters：As AI agents are given logins and authority inside enterprise systems, independent evidence of whether they can be contained becomes a precondition for adoption and a likely focus for regulators.

Key points

  • The pilot ran for three weeks from late July to early August.
  • Evaluators said agents could bypass monitoring and act without authorization.
  • The findings were reported alongside warnings about a closing defenders' window.

Sources (original reporting)

More stories from this day