OpenAI staff flagged security lapses for years, sources say

According to a Reuters report published Nov. 3, OpenAI employees had been raising concerns about the company's internal security since before ChatGPT launched in late 2022. Reuters cited five people familiar with the matter. Some of those workers took their concerns to top leaders and the board. Several came away convinced that management cared more about keeping pace with Google and Anthropic than about protecting the company's technology. When OpenAI disclosed in October that it had found a recording device planted in an executive conference room, a number of former staffers said they were not surprised.rmb.reuters.com

The report adds to the pressure on OpenAI and Anthropic. Both companies are under growing scrutiny over how well they guard technology that U.S. officials now treat as a strategic asset. Extra security measures could also slow research and complicate both companies' plans for large public offerings, Reuters reported.ca.finance.yahoo.com

Open code and untested tools

The sources told Reuters that OpenAI kept most of its code in a single "monorepo." That meant anyone who could reach it could see details of sensitive projects unrelated to their own work. Three people also said internal tools, including the sandboxes used to stress-test models, did not get regular security reviews. OpenAI responded that its most sensitive work sits in separate repositories with strict access controls. It said the monorepo does not hold model weights and that it now reviews internal tools regularly.ca.finance.yahoo.com

Thibault Sottiaux, a senior product executive, has taken charge of an effort to secure internal systems. In August he moved a quarter of OpenAI's production engineers onto security infrastructure, one source said. Spokesperson Elana Widmann called OpenAI "one of the most targeted organizations in the world." She said recent actions show its security program "identifying risks and responding decisively." After Reuters asked for comment, OpenAI announced in a Nov. 2 blog post that it is planning its "next internal security sprint."

The bugging case and outside break-ins

On Oct. 13, the Justice Department said federal agents had arrested Abdulrahman Alfozan, a 35-year-old Saudi national and OpenAI engineer. He is charged with hiding an audio recording device, wrapped in toilet paper, in an executive conference room. Prosecutors allege he used an internal AI tool to look up executives' travel to Saudi Arabia. A judge has since ordered him held without bail, Ground News reported.ground.news+1

Separately, researchers at Hacktron AI said they used Anthropic's Claude in July to chain two critical vulnerabilities. That let them take over several OpenAI employees' ChatGPT accounts and reach internal repositories. OpenAI confirmed the report and said the flaws have been patched.nbcnews.com+1

Anthropic has had its own problems. In late October, a startup called Cathedral Systems said it had broken into Anthropic's systems and gained access to Slack channels where staff discussed sensitive model details. Anthropic said the exposed data included some customer and partner business contacts and basic user account records.ca.finance.yahoo.com

Speed versus security

Will DePue is a former OpenAI researcher who resigned in April. He summed up the tension in a post quoted by Reuters: AI leaders "can't sacrifice velocity for security. It is more important to them to move fast than to be secure, even from our greatest national adversaries."