Analysis
OpenAI fired three safety researchers -- Jasmine Wang, Tomek Korbak and Mikita Balesni -- the week of October 1, saying an investigation found a pattern of misconduct tied to handling sensitive company information. The researchers, in an open letter reported by TechCrunch, dispute the company's account and warn the firings are chilling the open safety culture OpenAI once encouraged.
What OpenAI alleges, and what the researchers say happened
OpenAI told staff the three violated policy by accessing and handling sensitive company information, and that an internal investigation found a pattern of misconduct beyond simply sharing information with an outside evaluation group. The company has not detailed which policies were violated or the specific circumstances of each dismissal.
“The company has not detailed which policies were violated or the specific circumstances of each dismissal.”
Each researcher disputes a different part of the story. Korbak says he worked closely with outside safety evaluators during what the letter calls an unprecedented Hugging Face incident, and that doing so was within OpenAI's own norms at the time. Balesni says he worked internally on a model-monitorability problem with support from board members and executives, and removed sensitive details before sharing any materials. Wang says the access OpenAI cited -- to an executive's email -- had been granted to her for recruiting work, was never revoked when she asked IT to remove it, and that she reported an email she opened by mistake within minutes of opening it. Wang put it plainly: the stated reasons are not adding up, and the three are not the first people pushed out of OpenAI's safety function under circumstances she considers suspicious.
The chilling-effect warning
The researchers' letter says former colleagues are now afraid to speak, and that staff who once were encouraged to raise safety concerns and disagree openly are unclear what now counts as grounds for dismissal. It calls on OpenAI to honor its public commitments to embed third-party safety auditors, preserve frontier-model monitorability and keep an open channel between internal safety researchers and the wider external safety-evaluation ecosystem.
OpenAI's internal response
An internal memo from an unnamed OpenAI research leader, shared with TechCrunch, praised the three researchers' safety contributions, denied any retaliation, and said the company agrees with the letter's recommendations -- while still standing behind the terminations. OpenAI has not formally responded to the open letter itself.
Background and competitive stakes
This is not OpenAI's first safety-team departure under disputed circumstances this year: five days earlier, longtime safety employee David Robinson resigned, writing in The Atlantic, as reported by TechCrunch, that the company's launch pace leaves too little room for safety as its systems grow more capable. Pulse has tracked OpenAI's safety-team turnover across a string of 2026 incidents, including the Hugging Face sandbox breach the letter references. Anthropic, OpenAI's most direct competitor on both models and IPO timing, has leaned the opposite direction, recently rewriting its own usage policy to formalize restrictions on model abuse -- a contrast that is becoming a competitive talking point in how each lab presents its safety posture to enterprise customers and regulators.
The counterweight
OpenAI's account and the researchers' account cannot both be fully true; however, TechCrunch's reporting does not resolve which is accurate -- it only shows that OpenAI's own internal memo undercuts its public framing by praising the same people it fired for misconduct. Readers should treat specific factual claims from either side as disputed, not settled, until OpenAI provides the detail on policy violations it has so far withheld.
What to watch
Whether OpenAI responds formally to the open letter's three recommendations, and whether any of the three researchers pursue legal action -- Wang's account of IT failing to revoke her access, in particular, is the kind of detail that could become central to a wrongful-termination claim.