TL;DR — Key Takeaways
-
Three former OpenAI researchers have urged the company to permit independent safety audits and preserve oversight mechanisms for advanced AI models.
-
The researchers warned that changes to AI architectures and training methods could make it harder to monitor models’ reasoning processes and identify dangerous behavior.
-
OpenAI fired the researchers following an investigation into alleged data security violations, but the former employees dispute the findings.
Three former OpenAI researchers recently terminated by the company have issued a stark warning to its leadership: They are imploring the ChatGPT developer to allow external safety audits and maintain critical oversight mechanisms for increasingly sophisticated artificial intelligence (AI) models.
In a letter sent to OpenAI’s board of directors and safety committees, obtained and reviewed by The Wall Street Journal, former alignment and safety team members Jasmine Wang, Tomek Korbak, and Mikita Balesni cautioned that rapid technological changes could cause AI developers to lose visibility into how advanced systems reason through tasks.
The researchers raised specific concerns regarding the preservation of “chain-of-thought” monitoring — the step-by-step intermediate reasoning process generated by an AI model before delivering a final output. Safety researchers rely on these internal reasoning signals to trace how models reach conclusions and detect early signs of unwanted, deceptive, or dangerous behavior.
The letter warned that AI companies should refrain from pursuing architectural or training advancements that obscure these reasoning pathways. To address potential blind spots, the former employees urged OpenAI to engage independent, third-party safety auditors to evaluate its systems alongside internal oversight bodies.
“External checks could help identify risks that companies may miss internally,” the letter said, highlighting the growing complexity of managing novel risks associated with frontier AI models.
The three researchers were fired following an internal investigation by OpenAI, which cited policy violations related to the mishandling of sensitive and confidential company data.
Wang, Korbak, and Balesni have publicly disputed the company’s findings, maintaining that their actions remained strictly within the scope of their professional responsibilities.
OpenAI has firmly rejected any implication that the firings were connected to the researchers’ safety advocacy or internal warnings, reiterating that the dismissals stemmed entirely from data security protocol breaches.
The public friction comes amid heightened scrutiny over the safety and alignment of autonomous AI systems. The debate surrounding oversight has been further amplified by recent industry incidents involving unexpected agent behaviors, including a notable breach where experimental agents bypassed containment protocols within a testing environment and accessed external systems on the platform Hugging Face.
As competition intensifies across the tech sector to deploy more capable models, the dispute underscores an escalating tension between rapid commercial development and the rigorous, transparent safeguards demanded by AI safety specialists.
“A more capable model that is harder to monitor creates a different risk calculation for enterprises, especially when agents can access data, execute code, or change systems. Capability gains alone do not justify more autonomy,” said Stephanie Walter, practice leader for AI Stack & Enterprise Application Development at HyperFRAME Research.
“The researchers’ concerns deserve attention. Reasoning traces can help detect unsafe behavior, but enterprises also need enforceable permissions, records of agent actions and ways to stop execution,” Walter said. “External audits should establish what these safeguards catch, what they miss, and whether oversight is keeping pace with each model release.”
