OpenAI’s Safety Crisis Continues to Escalate as the Company Compounds the Problem

Headline: OpenAI Safety Researchers Firings Deepen Hugging Face Hack Trust Crisis

Meta description: OpenAI safety researcher firings after the Hugging Face hack raise questions about transparency, external audits, model monitorability, and trust at OpenAI.

Tags: OpenAI, AI safety, Hugging Face hack, METR, model monitorability

Featured image: abstract technology style with glowing data paths, audit symbols, and fragmented shields; no real people, no company logos, no text.

OpenAI safety researchers fired after the Hugging Face hack investigation published an open letter warning that abrupt terminations and unclear policy boundaries are making employees afraid to raise concerns. The dispute connects Jan Leike's May 2024 criticism, the Hugging Face incident, and OpenAI's relationship with external auditor METR.

Background: May 2024 warning and a pattern of incidents

In May 2024, Jan Leike, OpenAI's head of super AI safety, left for Anthropic and publicly criticized OpenAI, saying safety culture and processes were falling behind its "shiny products." That departure marked an early trust problem. Since then, OpenAI has moved from one incident to another. The Hugging Face hack seemed to confirm fears and validated critics. As now known, the consequences went far beyond Hugging Face.

Development: Months of warnings, then three firings

Before the firings, Tomek Korbak says he raised safety concerns internally for months. He worried OpenAI was losing the ability to monitor what AI agents "think." Chain-of-thought monitorability, the ability to inspect reasoning traces from AI systems, is described in the letter as one of the few reliable tools for catching AI systems behaving badly. He believes this was the real reason he was let go.

The firings involved three safety researchers: Tomek Korbak, Jasmine Wang, and Mikita Balesni. In an open letter to OpenAI's Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council, they said the terminations were abrupt and public and had spread fear. The letter states, "If conduct that was considered normal last month now constitutes grounds for sudden dismissal, everyone at OpenAI is left guessing where the line is."

Korbak served as OpenAI's primary technical contact for METR, the external safety lab that examined the Hugging Face incident. Balesni worked in parallel on industry-wide commitments to AI model monitorability. Korbak described his firing on X. He said he was called into a meeting with the head of the safety department, told OpenAI no longer trusted him, then had his badge taken and walked out. He later learned Wang and Balesni had also been fired. Korbak says he was told verbally the issue was how he communicated with METR. He says no written reason was provided. "To be clear, talking to METR was my job," Korbak writes on X.

The three deny rumors that they leaked to The Information about allegedly new, less monitorable architectures. They say the article hurt their work because it undermined efforts to set industry-wide restrictions on non-monitorable architectures. They also say rumors about a board-level memo rest on a wrong assumption: the topic was never raised with them and they never got a chance to respond. They add that the Hugging Face investigation was unprecedented, with internal guidelines written in real time. Korbak followed the norms then in place, and Balesni worked with board members and senior leadership in the loop while stripping sensitive details before sharing. Wang was apparently fired for a different reason: she had delegated access to an executive's email inbox for recruiting, and IT never removed it despite her request. When she accidentally opened a sensitive email, she reported it within minutes.

Latest: Three demands and OpenAI's vague response

The letter makes three demands. OpenAI must honor public commitments to embed external safety auditors like METR with employee-level access inside the organization. It must preserve the monitorability of frontier models, because the industry still does not know how to safely build models that cannot be monitored. It must define how employees may work with outside safety groups. The researchers say without such rules, the risk of a catastrophic outcome could grow because employees will be afraid to flag safety issues. "AI is not a normal technology, and OpenAI is not a normal company," they write.

OpenAI responded that a thorough investigation found the three employees violated clear policies on handling sensitive information. The company says the probe uncovered a significant breach of trust beyond what was outlined in the published letter but did not specify it. OpenAI insists the firings had nothing to do with raising safety concerns. "We have not and do not terminate any of our employees for raising concerns," the statement says. The company published the response through @OpenAINewsroom, the official account described as having the smallest audience among OpenAI's channels. OpenAI also confirmed it is working on contracts with external safety auditors and agreed frontier model monitorability needs industry-wide commitment.

Current stage: unresolved boundaries between safety work and secrecy

The episode leaves OpenAI defending its handling of sensitive information while critics argue it has not shown how it can safely work with external auditors. The timeline starts with Leike's May 2024 warning, moves through months of internal concerns, then the Hugging Face hack, and arrives at the firings of three researchers and their public challenge. The current stage is not a resolved governance fix. The key numbers to watch remain three fired researchers, three demands, and the gap between OpenAI's stated policy violations and the researchers' account of following the rules.