OpenAI stood by the firings in public on Friday and said outside safety testers are coming, but it gave no details on either.

OpenAI answered the three fired safety researchers publicly on Friday. In a note from its research leaders posted on X, the company said an investigation found that Jasmine Wang, Mikita Balesni and Tomek Korbak "violated clear policies on handling sensitive information." It added that the investigation "uncovered a significant breach of trust beyond what's outlined in the letter they published."

That second line is the new claim. The researchers' letter answered each accusation they say they were given: the September leak to The Information, Korbak's contact with METR and Wang's access to an executive's inbox. OpenAI is now saying the case against them goes further, but it hasn't said what that is. The company said it keeps employment matters private and doesn't think "a back and forth would be productive." It didn't respond to Wang's or Korbak's accounts point by point, and it didn't name the policies. Korbak has said he was told verbally that the problem was how he communicated with METR, and that nothing was put in writing.

The note also responds to the researchers' fear that OpenAI would cut back on outside testing. OpenAI said it is "actively finalizing contracts with third-party safety assessors" and will announce details "in the coming weeks," and that it is "committed to embedding external assessors." It didn't name METR or say what access the testers would get. Whether outside testers get the ongoing access Altman promised in September will depend on those contracts.

OpenAI repeated that the firings "were not about raising safety concerns or speaking out," and said it agrees that keeping AI reasoning readable "requires an industry-wide commitment, including from OpenAI." It made no new commitment on that point. Instead it pointed to research it has already published and to the GPT-6 Astra system card. The researchers had asked labs not to go ahead with work that makes monitoring harder. OpenAI's note doesn't say whether it will.