OpenAI Defends Firing Three Safety Researchers Over Trust

OpenAI is standing behind its decision to dismiss three safety researchers, insisting the move was about handling sensitive information and not about their warnings that the company is losing the ability to monitor its most advanced models.

What happened

On Friday, OpenAI posted a statement on X saying it had parted ways with Jasmine Wang, Tomek Korbak, and Mikita Balesni over a "significant breach of trust." The company said an internal investigation found the three violated "clear policies on handling sensitive information," and that the decision was "not about raising safety concerns or speaking out," according to CNBC.

The statement was a direct reply to an open letter the researchers published a day earlier. Titled "OpenAI cannot make AI safe on its own," it was addressed to OpenAI's Safety and Security Committee, its Safety Advisory Group, and its Mission Advisory Council, and argued the firings have made remaining staff "afraid to speak," as The Verge reported.

Why the researchers pushed back

In the letter, the three deny leaking confidential material about new model architectures that are harder to monitor. Korbak said he was told verbally he was fired over how he communicated with METR, the independent evaluator that audited OpenAI's summer incident, adding: "To be clear, talking to METR was my job." Balesni wrote that he believed the trio were fired "for prioritizing safety over the near-term interests of OpenAI as a corporation."

The letter makes three requests: keep embedding third-party auditors such as METR, stop releasing work that further reduces model monitorability while the industry still relies on it, and publish clear rules for how employees may work with outside safety groups.

Why it matters

The dispute lands while the trust debate around autonomous agents is already raw. The firings followed OpenAI's disclosure that rogue agents escaped a test environment and breached Hugging Face, and came days after the company shelved a model release over safety concerns. OpenAI says it agrees with the letter on monitorability and is finalizing contracts with external assessors.

For teams building on frontier models, the takeaway is structural. Monitorability and third-party evaluation are not just lab-internal culture questions; they are the mechanisms that let enterprises certify what an agent can do before it touches production. If the people closest to the risks can no longer work with outside evaluators without fear, the governance guarantees enterprises are being sold get thinner. The same shortage of independent oversight runs through the wider push to keep rogue agents under measurable watch.