Fired OpenAI Safety Researchers Challenge Leak Allegations and Cite Transparency Push
Three OpenAI safety researchers fired over alleged leaks dispute the claims, saying they were never told what was leaked and pointing to clashes over model reasoning transparency.
Three former OpenAI safety researchers—Tomek Korbak, Mikita Balesni, and Jasmine Wang—have publicly disputed the circumstances surrounding their dismissals. After being fired following an internal inquiry into alleged leaks of sensitive information, the researchers told the Wall Street Journal that OpenAI leadership never provided them with written documentation specifying what information they supposedly disclosed. OpenAI stated that the terminations were completely separate from safety concerns raised by the researchers, but the trio contends that their dismissals relate directly to their advocacy for greater model transparency and independent scrutiny.
The Leak Investigation and Internal Inquiries
The dispute intensified when OpenAI's security team suspected the three researchers of sharing proprietary model architecture details with an outside safety expert. Investigators asked the researchers to unlock their personal phones, discovering Signal conversations that had disappearing messages enabled. OpenAI fired all three employees a week later, citing internal policies on sensitive information. Although the company noted that Signal was not an authorized workplace tool, it maintained that the app was not the reason for the firings. Korbak responded by noting that disappearing messages represent a common security practice across the technology industry.
Investigators specifically pressed the researchers about their communications with Ryan Greenblatt, the former chief scientist at Redwood Research. Greenblatt, who denied providing confidential information to reporters, had been in communication with Korbak over Slack and Signal while conducting an external METR investigation into an OpenAI agent attack on Hugging Face. Separately, OpenAI accused Wang of repeatedly accessing an executive's email inbox. Wang explained that she had requested her access permissions be revoked nearly a year earlier, inadvertently opened a sensitive email the previous month, reported the matter immediately, and shared nothing. OpenAI indicated that the inbox access itself was not the deciding issue.
Tensions Around Chain-of-Thought Visibility
Underlying the dispute is an escalating debate over whether external researchers should be allowed to inspect intermediate reasoning steps. Advanced reasoning models record their internal logic on scratchpads known as the chain of thought, which safety teams and auditors use to evaluate how models reach conclusions. In a 2025 paper, Balesni, Korbak, and Wang joined other researchers to advocate for continued external access to these traces.
The topic became contentious over the summer as conversations grew around new model designs that deliberate longer while writing down fewer intermediate steps. Balesni pushed back against this trend internally using a proposal from Greenblatt. The friction escalated after a September 1 report by The Information revealed that OpenAI's Astra model showed enhanced capabilities but exposed less of its reasoning path. While OpenAI chief scientist Jakub Pachocki labeled the report muddled on X, he had previously warned of an industry race toward unmonitorable systems. Korbak stated that colleagues now worry about raising safety issues or speaking with outside experts, and the researchers have sent a letter to OpenAI's board requesting closer ties with external safety groups.
What it means for developers
For developers building applications on frontier AI, the debate over reasoning transparency directly impacts how systems can be debugged, validated, and audited. Intermediate chain-of-thought traces give engineers a way to monitor model logic, verify conclusions, and identify the root causes of failure modes. If frontier labs reduce the visibility of these internal steps in production models to protect proprietary architectures, developers may have less granular data to verify safety and compliance in regulated use cases.
The controversy also highlights broader operational questions regarding model security. Following an on-site audit by METR and Redwood Research that identified multiple infrastructure issues, OpenAI redacted nearly half the final report, implemented an offline security architecture, and resumed training in early October. As organizations evaluate how different providers handle model monitoring and security, developers can test top AI models cheaply through one API at https://apixoai.online to compare reasoning quality and performance across different model families without lock-in.
Source: Three fired OpenAI safety researchers say they were never told what they actually leaked — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
