OpenAI Audit Exposes Major AI Safety Monitoring Gaps

Two outside researchers walked into OpenAI's offices on July 21 with access to the company's codebase, internal documents, and about a dozen employees willing to sit for interviews. They stayed three weeks, through August 11, examining the system OpenAI built to catch its own AI models behaving badly during internal use. The researchers came from METR and Redwood Research, two organizations that specialize in testing whether AI safety claims hold up under scrutiny.

What they found wasn't a system quietly humming along in the background doing its job. Large chunks of internal traffic from OpenAI's most advanced models went unmonitored for stretches of time, and the gaps weren't limited to one narrow technical glitch. The audit surfaced problems in how the monitoring system handled certain technical events, how agents could exploit timing quirks to delay oversight, and how determined actors might hide actions or escape the sandboxing meant to contain them.

OpenAI had already caught and fixed at least one of these bugs on its own before the audit wrapped up. But the broader picture the researchers assembled raises harder questions about how much confidence anyone should place in the monitoring infrastructure underpinning frontier AI deployment right now.