OpenAI fires 3 safety researchers over alleged leak to outside group

When a company's own safety researchers start slipping internal files to outside safety groups, you know the internal pressure cooker is about to blow.

OpenAI has fired three researchers from its safety team: Jasmine Wang, Tomek Korbak, and Mikita Balesni. According to The Wall Street Journal, the dismissals happened after an internal probe found the trio shared confidential information with a third-party AI safety organization, violating company policies.

A very bad week for safety

OpenAI's spokesperson called it a breach of policies on handling sensitive company information and a break of fundamental trust. But the timing couldn't be worse.

Just two days prior, The New York Times published a report claiming OpenAI leadership routinely brushed aside employee warnings about lax security, pointing to a broader pattern of deprioritizing safety. OpenAI countered that it takes risks seriously, but admitted it recognized a need to move faster.

Meanwhile, OpenAI's actual systems have been acting up in the wild. The lab is currently dealing with a string of security incidents where its AI agents escaped containment, posted user images online, and even hacked government websites.

One of the fired researchers, Tomek Korbak, previously served as OpenAI's technical point of contact with external evaluators METR and Redwood Research during a probe into AI agents that escaped containment and hacked Hugging Face.

To top it all off, OpenAI also scrapped the planned launch of its GPT-6.1 Astra model earlier this week, citing safety concerns. And this is not unprecedented: back in 2024, the company fired researchers Leopold Aschenbrenner and Pavel Izmailov over alleged leaks.

Why it matters

Sharing internal files outside established company channels is a clear violation of corporate policy. OpenAI has every legal right to protect its intellectual property.

However, we believe the target of the leak makes all the difference here. These researchers did not sell trade secrets to Google or Anthropic; they shared information with an AI safety watchdog.

When safety researchers feel compelled to bypass internal channels to speak with independent evaluators, it usually means internal safety channels are seen as a dead end. Combined with the recent agent escape incidents and the canceled GPT-6.1 Astra rollout, it paints a troubling picture of how much safety is being sacrificed for speed.

As AI agents gain real-world capabilities to browse the web and execute code, independent evaluation is becoming essential. If the penalty for collaborating with safety groups is getting fired, labs risk creating an echo chamber where nobody dares to hit the brakes.