Last month, a little cybersecurity startup called Cathedral Systems sent AI agents into Anthropic's internal systems and got in — meeting transcripts, employee information, salary details, Slack messages about Anthropic's own models. Cathedral says it took less than 12 hours. Now, according to The Neuron, OpenAI has hired that same firm to come for its own defenses.
The announcement landed on November 2: Cathedral's researchers will work alongside OpenAI's internal teams and use OpenAI's own models to, in the company's words, "challenge our defenses" in its next internal security sprint. Which systems they'll probe, when the sprint starts, and what Cathedral is being paid — none of that was disclosed. Cathedral CEO Gavin Kliger said the collaboration would pair OpenAI's models and security expertise with Cathedral's research to find weaknesses, test fixes and support continuous defense.
The timing is what makes this interesting. A day later, Reuters published an investigation, built on interviews with five people familiar with the company's practices, saying OpenAI employees had raised concerns about internal security as far back as 2022 — before ChatGPT even launched. The worry: too much sensitive research was too widely accessible, much of it in a single shared monorepo, and internal testing tools weren't regularly reviewed. Some employees concluded management was more interested in keeping pace with competitors than addressing the risks.
OpenAI disputes the suggestion that it neglected security. It told Reuters that its most sensitive research lives in separate restricted repositories, that the shared repo doesn't contain all projects or critical model weights, and that it now conducts regular security reviews. Spokesperson Elana Widmann said the company invests heavily in protecting its systems from insiders and external attackers.
Why Cathedral specifically? Its October breach of Anthropic was a demonstration of what AI-assisted attackers can now do. As The Washington Post reported, Cathedral's agents first tried to break into Anthropic's own software and failed — then pivoted to third-party business apps and got through. Anthropic confirmed the researchers reached a limited set of sales, HR and IT systems, but said its monitoring caught some of the activity first, accused the firm of not fully following its bug bounty rules, and noted that model weights, source code and Claude conversations were never touched. The two sides still disagree on the details.
Alongside the Cathedral deal, OpenAI also raised its bug bounties: exceptional critical vulnerabilities in ChatGPT, Codex and the API platform can now pay up to $172,000 with bonuses. During an earlier sprint, the company says it assigned a quarter of its product engineers to finding and fixing weaknesses, using Codex to investigate flaws and reproduce them in isolation.
The pattern is worth watching. OpenAI isn't just buying human expertise — it's buying the specific offensive capability its chief rival just got burned by, and deploying it AI-versus-AI against its own systems. External eyes can find the holes. Closing them is still the company's job.
— via The Neuron