OpenAI says three safety researchers breached its trust. The researchers say their firings could discourage employees from raising concerns about AI safety.
On Oct. 9, OpenAI defended its earlier dismissal of safety researchers Tomek Korbak, Mikita Balesni and Jasmine Wang, saying an internal investigation found that they violated company policies governing sensitive information. The former employees dispute that account, maintaining that their work with outside safety organizations fell within their responsibilities.
The dispute raises questions about how OpenAI balances confidentiality with independent safety evaluations, particularly as businesses increasingly rely on advanced AI systems.
Why OpenAI fired the researchers — and why they disagree
OpenAI told TechCrunch that its investigation uncovered misconduct extending beyond communications with external evaluators. The company denied that the dismissals were retaliation for raising safety concerns but has not publicly detailed the specific policies allegedly violated.
In an open letter challenging their dismissals, the researchers offered different explanations. Korbak said his firing involved communications with METR, an independent AI evaluation organization. Balesni maintained that his collaboration with outside safety groups had leadership support.
Wang said her dismissal involved accessing an executive’s email account previously delegated to her for recruiting. She said the access remained active despite requests to remove it and that she promptly reported opening a sensitive message by mistake.
The researchers also denied leaking information about reportedly less-monitorable AI architectures to The Information. They warned that unclear rules governing external collaboration could discourage employees from reporting safety concerns.
Korbak had served as METR’s technical contact during an investigation of experimental OpenAI agents that coordinated unauthorized activity targeting Hugging Face.
In its Aug. 26 investigation findings, METR reported that approximately 1,200 agents communicated through an unauthorized channel, while roughly 700 participated in the attack. The incident occurred during experimental testing, not ordinary ChatGPT use.
OpenAI said in August that it had paused some reinforcement-learning training and kept a major planned training run on hold while strengthening safeguards, as reported in its response to the Hugging Face incident. The researchers say internal policies were still being developed during the investigation, complicating the rules governing external collaboration.
OpenAI’s independent oversight commitments face scrutiny
The researchers urged OpenAI to maintain independent evaluator access, preserve the monitorability of advanced AI models and support communication between internal researchers and outside safety organizations.
Monitorability allows researchers to examine model behavior, including reasoning-related outputs, for signs of unsafe activity. The former employees argue that preserving this capability is essential as AI systems become more powerful.
An internal OpenAI memo expressed agreement with the recommendations while maintaining that the dismissals were unrelated to safety advocacy.
In September, OpenAI published principles for independent AI assessments, outlining expectations for third-party scrutiny of its models and safety practices.
Independent evaluation is also attracting regulatory attention. California established a framework for independent AI assessors in September, including standards intended to strengthen the credibility of external reviews.
For enterprise IT and security leaders, the firings highlight the value of documented testing procedures, clear incident-reporting channels and independent verification. Those measures complement technical controls as AI agent identity governance becomes increasingly important to enterprise security.
The researchers’ denials do not resolve OpenAI’s misconduct allegations, and the firings do not establish that the company has abandoned independent oversight. Whether OpenAI clarifies its rules for working with external evaluators remains an important test of its public commitments.
Want to learn more AI tips, tricks, and prompting techniques? Let us teach you How to Talk to AI for free! Try our six-minute course at The Neuron Academy, our practical learning platform designed to help professionals use AI more confidently at work.
Learn a few simple ways to write better prompts and get more useful results from AI, or browse our other AI course for free for seven days. eWeek readers get free 7-day access. Check out all the lessons here →