Security researchers used Claude to help them hack into OpenAI

A team of three independent security researchers at Hacktron says it used Anthropic’s Claude Opus 4.8 and 5 to break into OpenAI employee accounts in less than 72 hours, according to the Wall Street Journal. That is a fast turn for a job that, on paper, sounds like the sort of thing a company would have a few more guardrails around.

The prize was OpenAI’s GitHub repository, Monorepo. The Journal says its sources described that repository as holding OpenAI’s “algorithmic secrets.” The researchers say they did not actually get into the internal code there themselves, but they did enough to show they had access.

Their proof was a pull request sent from an employee’s Codex account. That is a neat little flex, and a reminder that proof of access can be more persuasive than a pile of screenshots.

The route in went through Discourse, the third-party service that hosts OpenAI’s community forum. That detail matters because breaches often don’t begin at the shiny, heavily defended center. They start at the awkward edges, where a vendor, a login flow, or some forgotten account opens the door a bit wider than anyone expected.