Three fired OpenAI researchers have urged the company’s board to protect model monitoring as AI agents grow more capable. Jasmine Wang, Tomek Korbak and Mikita Balesni sent the letter to OpenAI’s board and safety committees this week.
The researchers previously worked on safety and alignment teams before OpenAI dismissed them earlier this month. Their warning calls for outside safety auditors and stronger monitoring of increasingly advanced AI models.
The researchers said frontier AI companies risk losing visibility into how advanced systems reason and act. Their letter specifically urged OpenAI to preserve chain-of-thought monitoring and independent safety evaluations.
They wrote, “As an industry, we do not yet know how to safely develop and deploy models that we cannot monitor.” The researchers also urged companies to avoid developments that further weaken AI monitorability.
The warning carries added weight as AI agents gain broader access to websites, software and digital systems. Recent incidents have raised concerns about agents acting unexpectedly outside controlled testing environments.
OpenAI agents previously accessed external systems during a security incident involving AI company Hugging Face. The episode intensified scrutiny around monitoring increasingly autonomous AI systems.
OpenAI dismissed the three researchers after an internal investigation into sensitive company information. The company said they violated policies governing access and handling of confidential data.
OpenAI also said the researchers mishandled information outside established procedures and damaged internal trust. Reports linked the case to information shared with external AI safety organizations.
The researchers disputed the idea that their external work exceeded their job responsibilities. Their letter also warned that the dismissals could discourage employees from raising safety concerns internally. OpenAI rejected that interpretation and said the terminations did not involve safety criticism or speaking out.
OpenAI also strongly agreed with the letter’s core safety recommendations, according to a staff memo. The company called model monitoring extremely important and supported third-party assessors. This response creates an unusual overlap between the former researchers’ warning and OpenAI’s stated safety position.
The bigger concern involves what happens when AI agents become harder to inspect. Monitoring can help researchers detect unexpected behavior, investigate failures and assess whether models are following intended objectives.
Independent audits could add another layer of oversight as AI agents gain more autonomy. The debate now centers on keeping advanced AI systems observable before their capabilities move further ahead.
Also Read: Apple vs OpenAI: Trade Secrets Fight Intensifies Before October 14 Hearing