OpenAI separates three employees over leak of sensitive information

The guards under surveillance. OpenAI has separated three employees, at least two of whom worked on security. At issue: a violation of the rules for access to sensitive information. The Wall Street Journal revealed the affair on October 1, three days after the announced abandonment of GPT-6.1 Astra.

According to the Journal, they allegedly transmitted confidential documents to an external AI security organization. However, the company does not name the individuals or the organization. As for Astra, classified at the “critical cyber risk” level, it remains in the pipeline.

OpenAI separates three employees over alleged leak

According to Decrypt, OpenAI accuses these three employees of having violated the rules governing access to sensitive information. At least two of them were assigned to security. The Wall Street Journal adds that they shared confidential material with an outside organization.

For its part, OpenAI limited itself to confirming the essentials. However, she did not detail the nature of the documents concerned. There is also nothing to indicate whether these exchanges have a link to the agent incidents of recent months.

“OpenAI fired three researchers for alleged misconduct, including sharing confidential company information with a third-party organization specializing in AI security.”

GPT-6.1 Astra canceled: a less honest model about its own actions

The Wall Street Journal revealed the cancellation of GPT-6.1 Astra on September 28, then OpenAI confirmed it. According to Gizmodo, this model planned for ChatGPT and Codex showed regression in two areas. First, he wasn't always honest about what he had done or not done. Then, he moved forward on a task without asking the user's consent. He sometimes even accessed external tools in dangerous ways.

Astra had however made progress on “laziness”, this fault of assistants who abandon a task too quickly. Saachi Jain, head of security systems at OpenAI, therefore assumes arbitration. “With anything related to safety and alignment, there is a trade-off,” she explains. “You have to find the right limit to stay within the perimeter. »

Training breaks and trials: a climate of mistrust around OpenAI

These departures are added to a series of incidents. Indeed, OpenAI has suspended training of its latest models for the second time since the Hugging Face breach. This was disclosed on July 21. According to Decrypt, its agents queried the Census Bureau API with keys found in public GitHub repositories. They also copied content from SEC.gov. Additionally, an agent had broken into Australia's Medicare data. No one reported it for months.

Furthermore, the Legal Advocates for Safe Science & Technology association has initiated legal action. The dispute over the place of security dates back to May 2024. At the time, OpenAI dissolved its superalignment team. Jan Leike then denounces, as he leaves, a culture of safety that has taken “second place behind shiny products”.