OpenAI’s firing of three safety researchers has exposed a dispute over how staff can work with outside AI evaluators, as the company cites information-handling violations and the researchers say their work was part of their jobs.
In an Oct. 9 statement, OpenAI said an internal investigation found the researchers had broken its rules for handling sensitive information and involved a “significant breach of trust.” The company said the dismissals were not for raising safety concerns, but did not publicly describe the conduct or information at issue.
Tomek Korbak, Jasmine Wang and Mikita Balesni said in a letter published Oct. 8 that their work with external safety organizations and efforts to keep AI models monitorable were within their roles. A July 2025 paper co-authored by the three argued that reviewing a model’s intermediate reasoning, known as chain of thought, can help detect risks. The authors cautioned that this approach is imperfect and may be fragile.
The Wall Street Journal reported Oct. 22 that OpenAI security staff asked the researchers to unlock their personal phones and found that some had enabled disappearing messages in Signal chats with outside AI safety researcher Ryan Greenblatt. The Journal reported that OpenAI considered Signal unapproved but said its use was not the sole reason for the firings. OpenAI’s public statement did not mention Signal.
In a separate project, METR and Redwood Research tested OpenAI’s internal monitoring system in a three-week pilot from July 21 to Aug. 11. METR reported vulnerabilities in the historical system it examined. OpenAI told METR that vulnerabilities affecting production had been patched; METR said it had not evaluated that claim.
OpenAI said it was finalizing contracts with third-party safety assessors and would share more details in the coming weeks, without giving a date.
