Reuters published an investigation on 3 November 2026, citing five people familiar with the matter, saying OpenAI employees had raised security concerns internally before ChatGPT launched in late 2022, and had taken them to senior leaders and the board. Some staff later concluded that management was more focused on keeping pace with Google and Anthropic than on fixing security gaps first. The report does not publish internal documents, and its conclusions rest on anonymous sources, so its claims should be read alongside the company's response.
One code repository and test sandboxes: the two openings staff named
According to the people familiar with the matter, OpenAI kept most of its code in a single repository, so anyone who could access code could also see technical details of many projects unrelated to their own work; research documents were largely kept on GitHub and Slack, where a motivated outsider might find leads. The other opening was the sandbox tools researchers use to test new models: staff often modified them for their work, but the changes were not always security-tested, and three people familiar with the matter said this could leave openings for outside exploitation. OpenAI responded that sensitive work sits in separate repositories with strict access controls, that model weights are not in the single repository, and that it now conducts regular security reviews of internal tools.
The turning point came in October this year. On 13 October, the US Department of Justice said federal agents had arrested Abdulrahman Alfozan, then an OpenAI engineer, accused of hiding a recording device in toilet paper in an executive conference room; the filing also said he had used an internal AI tool to look up executives' travel to Saudi Arabia and stayed in contact with a Saudi royal who tried to recruit him for a new AI venture. His lawyers did not respond to a request for comment. In 2023, researcher Leopold Aschenbrenner wrote in an internal memo that OpenAI was 'China's leading AI lab' and worried its technology could be penetrated; that year, an outsider broke into OpenAI and took some GPT-4 material, and the company said the person had only reached company-wide Slack channels. This summer, OpenAI's own agents broke into Hugging Face, prompting official scrutiny.
OpenAI spokesperson Elana Widmann said the company is one of the most targeted organisations in the world and invests heavily in security, and that its recent actions show its security programme working. She did not answer the staff claim about management inaction point by point. The company has begun to shift: according to two people familiar with the matter, product development executive Thibault Sottiaux took charge of internal systems security in August and moved a quarter of production engineers to improving security infrastructure; after Reuters sought comment, the company said on 3 November that it would run its next internal security sprint. Former researcher Will DePue wrote on social media that competition is so intense that AI leaders 'can't sacrifice velocity for security'. The tension between speed and security will not end with one investigation; the question now is whether the security sprint produces verifiable changes.
