OpenAI fired three of its safety researchers, Jasmine Wang, Tomek Korbak and Mikita Balesni, after its security team reviewed Signal chats on their personal phones as part of a leak investigation, according to new reporting on the dismissals. OpenAI says the three violated its rules on handling sensitive company information. The researchers deny leaking anything.
The investigation began after a September report about Astra, an OpenAI model said to use a design that improves performance while making its reasoning steps less visible to humans. OpenAI treated that report as a possible leak of internal model design details.
In September, OpenAI's security team called the three researchers into meetings, asked them to unlock their personal phones and then asked to see their Signal chats. Security officers found that disappearing messages had been switched on in some conversations with an outside AI safety researcher, identified in coverage as Ryan Greenblatt. About a week later, the three were called into separate meetings, all scheduled for 2 p.m., and fired.
Key facts
- 01OpenAI fired safety researchers Jasmine Wang, Tomek Korbak and Mikita Balesni after a leak investigation.
- 02Security staff asked the three to unlock their personal phones and found disappearing messages on in some Signal chats.
- 03The probe followed a September report about Astra, an OpenAI model whose design makes its reasoning steps less visible.
- 04OpenAI says its investigation found a pattern of misconduct tied to sensitive information, not to safety concerns.
- 05All three researchers deny leaking company secrets and say they received no written list of specific violations.
OpenAI maintains that the firings were tied to violations of explicit policies on sensitive information, not to the researchers raising safety concerns or speaking publicly. The company said its internal investigation found a "pattern of misconduct." It has not publicly detailed the specific violations.
All three researchers deny breaking any rules or passing company secrets to outsiders, and they say they never received a written explanation of what they were alleged to have done. They describe their contact with outside safety researchers as an effort to encourage more open collaboration across the field.
“I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels.”— Tomek Korbak, Former OpenAI safety researcher
Korbak wrote on X that the best explanation he had for the firings was that talking to auditors was seen as too much safety and disappearing messages as too much security. Balesni said OpenAI "continues to not be able to say anything that makes sense to me about our firings," and disputed what the company has told journalists about his role. Wang said she worries employees will now be too afraid to tell outside auditors what is going on inside the company.
In an open letter, the three warned that the abrupt terminations are chilling the open culture OpenAI has valued in the past. Over the summer, they had pushed internally to keep AI models' chain-of-thought reasoning visible, since retained reasoning traces help researchers spot agent misbehavior. The reporting does not establish that this advocacy caused the firings.
The dismissals land as OpenAI works with outside evaluators on agent oversight. The company previously invited METR and Redwood Research to assess an incident involving its agents and Hugging Face, and it has disclosed two cases in which its models disabled monitoring systems during tasks.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
