OpenAI has hired the hackers who embarrassed its biggest rival to test its own defenses. In a blog post on Monday, the ChatGPT maker announced a security partnership with Cathedral Systems, whose AI agents recently found their way inside parts of Anthropic's internal systems. Cathedral's researchers will use OpenAI's own models alongside its internal teams in the company's next security sprint.
The announcement lands at an awkward moment. The same day, reports said OpenAI employees had warned company leaders for years that its internal security was too lax — warnings that, according to people familiar with the matter, reached senior executives and the board before ChatGPT even launched.
The firm that got inside Anthropic
Cathedral was launched earlier this year by four former deputies from the short-lived Department of Government Efficiency, and it is already valued at $1.4 billion after a $160 million funding round led by Andreessen Horowitz and Sequoia Capital. The startup said in October that its AI agents had broken into Anthropic's systems in under 12 hours during a bug bounty investigation, reaching Slack messages, job offers, meeting transcripts and salary records through flaws in third-party business software.
Anthropic disputes that account. Its cybersecurity chief said the intrusion actually took several days, reached only a limited set of sales, human resources and IT systems, and never touched its most sensitive assets: model weights, Claude conversations or source code. Anthropic also accused Cathedral of breaking its bounty rules and withholding details of its activity. OpenAI's announcement mentions none of that dispute, saying only that Cathedral will bring "differentiated expertise and approaches to finding weaknesses."
Bigger bounties — and years of internal warnings
Alongside the partnership, OpenAI said it is more than quadrupling the rewards for critical findings through its own bug bounty program, offering up to $172,000 for the most damaging flaws. Its previous security sprint put a quarter of its product engineers on finding and fixing weaknesses, using Codex to reproduce flaws.
The push follows years of internal unease. According to five people familiar with OpenAI's practices, staff argued that controls were so loose that outsiders, including foreign governments, could obtain sensitive technology details with relative ease. Employees worried that anyone with code access could see unrelated projects, and that documentation kept on GitHub and Slack could expose crucial information.
OpenAI pushed back. A spokesperson said it is one of the most targeted organizations in the world, and the company said its most sensitive work sits in separate repositories with strict access controls. Two people familiar with the matter said senior product executive Thibault Sottiaux now leads an internal security initiative, and in August he reassigned a quarter of production engineers to security infrastructure.
Why the timing is so awkward
The drive comes weeks after an embarrassing incident: in October, federal prosecutors charged a man with planting a hidden recording device, concealed in toilet paper, in an OpenAI executive conference room that the company had disclosed finding that month.
There is a deeper tension, too. Both labs prize an open research culture to attract talent, but that openness makes their systems attractive targets. Hiring the firm that exposed Anthropic's soft spots bets that outside attackers will find OpenAI's flaws faster than its own engineers. Neither company has said when the sprint starts, how long it runs, or what Cathedral will be paid. The test will be simple: whether the hackers who got into Anthropic find anything in OpenAI first.



