Former OpenAI researchers raise concerns about outside safety audits

In an interview published October 23, 2026, former OpenAI safety researcher Mikita Balesni said negotiations over continued embedded auditing were not going well, in his view. OpenAI had published its principles for third-party safety assessments on September 22, laying out four areas for outside review and rules for how it should work.

What OpenAI says outside assessments should cover

OpenAI names four priorities: safety cases, critical safeguards, capability and alignment evaluations, and investigations of critical misalignment incidents. The company’s published assessment principles describe how it says outside reviews should be organized, conducted and shared.

The framework calls for assessors and OpenAI to agree on the scope and claims being assessed, with access proportionate to that work. It also calls for transparent methods, relevant expertise and independence, and protections for security and confidential information. If assessors find significant risks beyond the agreed scope, OpenAI says there should be a way to raise them.

OpenAI says assessments should produce actionable findings and allow time for remediation. Reports, it says, should be shared as openly as possible, with explanations for redactions. The company describes assessments that may last weeks or several months, depending on the question, and says this work is generally longer-term and not tied to a product launch.

What Korbak and Balesni said about outside access

For the summer 2026 incident involving OpenAI agents and Hugging Face, researchers from METR and Redwood Research were invited to investigate. In the October 23 interview, former OpenAI researcher Tomek Korbak said METR initially expected to spend two days on site. As the incident’s scale became clearer, he said, the assessors were invited back for a second and third visit.

Korbak described the investigation as restricted, while saying he felt OpenAI was broadly acting in good faith. He also said it was difficult to know in advance which questions an auditor should ask and what to inspect. His account reflects his assessment of that investigation.

Balesni described negotiations over continued embedded auditing and employee-like access as not going well, in his opinion. OpenAI’s September principles describe proportionate access, including indirect or privacy-preserving approaches when direct access is impractical. Balesni’s account concerns discussions over a particular arrangement; the principles set out the company’s intended approach.

El Mundo’s account of The Wall Street Journal’s investigation reported that OpenAI announced a pilot involving METR and Redwood Research. It also reported that METR said OpenAI censored some results it could share.

The researchers’ dismissals remain disputed

OpenAI dismissed Balesni, Korbak and Jasmine Wang earlier in October 2026. In reporting published by NPR and El Mundo, OpenAI attributed the dismissals to mishandling sensitive information after an internal investigation. The researchers disputed the company’s explanation.

The dispute over the dismissals is covered in NeoTeo’s earlier report on the researchers’ competing accounts.