OpenAI put the AI race ahead of security, employees tell Reuters

An audio recorder hidden in toilet paper in an OpenAI executive meeting room: on October 13, US authorities announced the arrest of a company engineer, accused of planting it there (we covered the story). Former employees were not surprised, news agency Reuters reports in an investigation published on November 3.

According to five people familiar with the matter, employees have been warning management and board members for years that operational security was too lax, some even before the release of ChatGPT in late 2022. Several concluded that management preferred to keep pace with Google and Anthropic rather than lock down its systems. These employees believed outside actors, including foreign governments, could obtain sensitive information with relative ease.

What Reuters' sources describe

• An almost single code repository. Most of OpenAI's code was in one repository. Anyone with access could view the details of many projects, including sensitive ones unrelated to their work. AI agents now make this kind of digging easier, Reuters notes.
• Research documents stored on GitHub and Slack, which worried some employees.
• Loosely controlled testing environments. According to three sources, the “sandboxes” used to test models were not regularly audited, and employees' changes were not always reviewed.

The microphone case illustrates the risk: according to a court document cited by Reuters, the engineer allegedly used an internal AI tool to find documents about executives' trips to Saudi Arabia. Reuters also recalls that in July 2026, three researchers from Hacktron AI used, according to the company, an Anthropic model to break into ChatGPT accounts of OpenAI employees. They reached some internal systems, including this repository. OpenAI says it thanked them and took action.

OpenAI disputes the claims, but is reorganizing

Its spokesperson, Elana Widmann, describes OpenAI as “one of the most targeted organizations in the world,” investing heavily against insider threats, cybercriminals and states. According to the company, its most sensitive work is kept in separate, strictly access-controlled repositories, and the shared repository contains neither all the code nor the model weights, meaning their parameters. She says it now regularly audits its internal tools, under the leadership of its head of security, Dane Stuckey, who joined in October 2024.

Two sources nonetheless describe a shift. Thibault Sottiaux, a product executive, has taken charge of an internal security effort. In August, according to one of them, he reassigned a quarter of production engineers to security infrastructure.

On Monday, November 2, after questions from Reuters, OpenAI announced its next internal security “sprint.” It is turning to Cathedral, a startup founded by four former members of DOGE, the government spending-cut unit launched by Elon Musk. Its teams “will use OpenAI's models alongside our teams to put our defenses to the test,” the company writes. It also says it is more than quadrupling the bounties paid for critical vulnerabilities reported to it.

A partner under scrutiny itself

The choice is raising eyebrows. In late October, Cathedral claimed to have broken into Anthropic through its bug bounty program (read our article). The startup told the Washington Post it did so in under twelve hours and gained access to Slack messages, job postings and meeting transcripts. Logan Graham, head of cybersecurity at Anthropic, speaks of “several days” and calls the method “unusual and out of scope.” According to him, model weights and conversations with Claude were not affected. Anthropic acknowledges access to business contact details of customers and partners and to basic account data, according to Reuters.

What your company can take from this

The criticisms leveled at OpenAI apply to many companies: overly broad permissions on code repositories, internal tools modified without review, and AI assistants plugged into every document. Before connecting an agent to your shared storage, clean up access rights: the agent will find everything its user can read.

The dilemma goes beyond OpenAI. In such a tight race, AI leaders “cannot trade speed for security,” Will DePue, a researcher who left the company in April, summed up on X. It remains to be seen whether OpenAI will make Cathedral's test results public.