OpenAI plans security push after years of staff warnings

OpenAI plans another security overhaul after employees warned that broad internal access could expose sensitive AI research to outsiders.

The company announced another round of security work in a blog post on Monday. Five people acquainted with its operations described employee warnings dating to before ChatGPT’s introduction in 2022, with concerns reaching senior management and directors.

OpenAI spokesperson Elana Widmann said the company devotes considerable resources to defending against employees acting improperly, criminal hackers and government-backed attackers. She characterized recent interventions as evidence that its security program identifies and addresses threats.

Engineering resources shift toward security

Thibault Sottiaux, the executive responsible for product development, recently assumed oversight of an effort to strengthen internal systems, according to two people familiar with the work. One said he redirected 25% of the production engineering workforce toward security infrastructure in August.

That allocation puts engineering capacity at the center of a dispute over research speed and protection. Some employees who raised concerns concluded that leaders prioritized competition with Google and Anthropic, according to the five people; that assessment represents the workers’ interpretation, not an established management motive.

The tension extends beyond a single company. OpenAI and Anthropic have cultivated collaborative research environments to recruit scientists and encourage discoveries, drawing on an academic tradition of sharing findings; tighter restrictions could complicate that approach.

Access controls face competing accounts

The five people described a shared code repository that allowed employees with access to inspect projects beyond their assignments. They also said research records on GitHub and Slack created concerns about how easily someone could locate valuable information.

OpenAI disputed the breadth of that exposure, saying particularly sensitive projects reside in separate repositories protected by restrictive permissions. It said the shared repository excludes model weights, the numerical parameters governing model behavior, and does not hold its entire codebase.

Three people familiar with internal practices said security checks on research tools had previously been irregular, including after employees modified testing environments. OpenAI said it now reviews internal tools regularly under chief information security officer Dane Stuckey, who arrived in October 2024.

Recording charge adds insider scrutiny

The Justice Department announced on October 13 that federal authorities had arrested OpenAI engineer Abdulrahman Alfozan and charged him over an alleged concealed recorder in a room used for executive meetings. A court filing also alleged that he searched internal documents concerning executive travel to Saudi Arabia and meetings with Saudi ministers.

According to that filing, Alfozan maintained contact with a Saudi royal seeking to recruit him to an AI venture. His lawyers did not provide a response for the supplied report, and the material contains no adjudication of the recording-device allegation.

External researchers have separately described digital access failures. Cybersecurity company Hacktron AI said its researchers used an Anthropic model in July to compromise employee ChatGPT accounts and reach internal resources, including the shared repository; Widmann said OpenAI thanked them and acted to reduce recurrence risks.

Those accounts leave a specific unresolved issue: how far the newer restrictions address the access paths employees and researchers described. OpenAI has outlined additional security work, but the supplied material provides neither a completion schedule nor an independent assessment of its effectiveness.

If stronger controls limit unauthorized access without obstructing legitimate collaboration, OpenAI could protect research while preserving development capacity. If restrictions slow experimentation, the company and competing laboratories would face a harder trade-off between safeguarding technology and advancing it, with implications for national efforts to retain AI capabilities.