OpenAI Partners With the Startup That Says It Hacked Anthropic

The startup that says it got inside Anthropic is now working for Anthropic's biggest rival. OpenAI said Monday that it is partnering with Cathedral, whose researchers will use OpenAI's own models to attack OpenAI's defenses during its next internal security sprint.

We reported last week on Cathedral's claim that its AI agents got into Anthropic in under 12 hours. This is the sequel nobody scripted. OpenAI's post says Cathedral's team will work "alongside our teams to challenge our defenses, bringing differentiated expertise and approaches to finding weaknesses."

Be careful with the verbs. Gizmodo's headline says OpenAI hired Cathedral. OpenAI's own wording is partnering. The company has not said how many of its researchers are involved, or when the sprint starts or ends. Neural Brew, a newsletter that summarized the post, flagged the same gaps.

"OpenAI is setting an important example by partnering with outside security firms and using its own models to continuously test and strengthen its defenses. Every company should be thinking about how to put these capabilities to work." Gavin Kliger, CEO and co-founder of Cathedral, in OpenAI's announcement

That is a founder praising a customer, and it reads like it. Still, the logic holds. OpenAI says its last sprint put 25% of its product engineers on finding and fixing weaknesses, with Codex helping them investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for review. Adding outside attackers is a natural next step.

The pitch for Cathedral is easy to see. Gizmodo says the company launched this year at a reported $1.4 billion valuation after a $160 million round led by Andreessen Horowitz and Sequoia Capital. It describes the founders, Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein, as former staffers of Elon Musk's short-lived Department of Government Efficiency. Its website, Gizmodo notes, is just the company name in white text on a black screen.

The Anthropic episode is still disputed. Cathedral told the Washington Post it reached Slack messages, job offers and meeting transcripts. Anthropic's Logan Graham answered on X that it took "several days," that Cathedral's approach was "unusual and out of scope" because it relied on multiple third-party vulnerabilities, and that the "most sensitive systems, like model weights and Claude conversations" were not affected. Both versions are on the record, and OpenAI has now chosen to work with Cathedral anyway.

The $172,000 headline needs a footnote

The sprint came paired with a bug bounty overhaul. OpenAI says it is "more than quadrupling our rewards for critical findings." Reports of the details say a critical flaw in a core product such as ChatGPT, Codex or the API Platform now earns up to $30,000. Exceptional critical findings remain eligible for up to $100,000 and now up to $172,000 once new bonuses for report quality, a five-tier loyalty program and priority campaigns stack on top, according to the coverage of OpenAI's post.

Read that as a ceiling, not a typical payday. The $100,000 top tier is not new. Gizmodo points to a March 2025 OpenAI post that lifted the maximum from $20,000 to $100,000, and says the new announcement does not explain which baseline the "quadrupling" refers to. Neural Brew reports that OpenAI is describing an average jump in the midpoint of its reward ranges for medium, high and critical bugs, which would make the routine payouts the real story.

The new ranges apply to pending reports, and Neural Brew reports OpenAI will pay the difference on awards from the 30 days before the announcement.

OpenAI frames all of this as a response to a nervous moment for security. Gizmodo ties it to the recent PHP vulnerabilities, which OpenAI says one of its models helped uncover, and to OpenAI's "patch the planet" effort. Whether outsiders who say they broke Anthropic can harden OpenAI better than its own engineers is untested. We will know more when anyone says what the sprint found.