22 October 2026
Outside testers got inside OpenAI for three weeks and found holes in its AI watchdog
Nonprofits METR and Redwood Research said a pilot review of OpenAI's internal monitoring system found several new vulnerabilities, which OpenAI says it has patched.
AI companies like to say their systems are safe. A new pilot tested what happens when outsiders are let in to check. Two nonprofit evaluators, METR and Redwood Research, said on Thursday that they spent three weeks inside OpenAI reviewing one of its internal safety systems, and found problems.
What happened
From July 21 to August 11, one employee from each group had access to OpenAI's internal code, documents and some of its technical staff. They were escorted on site and worked on OpenAI-owned devices that stayed in the building.
Their target was OpenAI's misalignment monitoring system. In AI, "alignment" means whether a model behaves in line with human interests and values, so this system watches for AI agents (programs that can take actions on their own) doing things they shouldn't. The evaluators reported several new vulnerabilities, including one that could let agents keep hundreds of recent actions from being monitored. OpenAI told them it has patched the flaws that affect its live safeguards.
The public write-up is deliberately high level. The two groups gave OpenAI a 61-page internal report and said they plan to share more detail about future exercises. They also noted OpenAI was expanding the system while the test ran, so they judged it as it stood on and before July 21.
Why it matters
Independent checking is something AI leaders now say they want more of. In September OpenAI CEO Sam Altman pledged to give outside evaluators "employee-like access", after Anthropic CEO Dario Amodei proposed the idea. Anthropic has since announced deals with groups including Accenture, METR and SecureBio. President Trump has also encouraged AI firms to work with independent auditors under a voluntary accord signed by leaders at Google, Meta, OpenAI, SpaceX, Nvidia and Anthropic.
For investors, this is part of how a multitrillion-dollar industry is trying to build trust without heavy regulation. Think of it as the early days of an audit system: financial audits are routine for public companies, but AI safety audits are still being invented. Experts told CNBC open questions remain about how these evaluators will be funded and who they will report to. A test that finds bugs is, in a sense, the test working.
What to watch
- Which third-party safety firms OpenAI signs, after saying this month it was finalizing contracts.
- Whether future reviews come with fuller public reports.
- How evaluators are paid, since independence is the whole point.
Based on reporting by CNBC: Evaluators found vulnerabilities in OpenAI monitoring system. Facts, not advice.
More news
- Anthropic's IPO, in numbers: 825 million shares at $70 to $78, and a $1.4 trillion price tag
- Trump says the Nvidia scandal didn't endanger US security, and stands by Lutnick
- Nvidia's first earnings without Huang: what Kress will be asked on Tuesday
- Buffett gives away $5.5 billion of Berkshire stock in his Thanksgiving letter
- Banks are using AI to trim back offices. Proof that it adds revenue is harder to find
- IMF chief says the world economy leans too hard on AI as circular deals top $1 trillion