OpenAI lines up Cathedral Systems to test its defenses with its own models
On November 2, OpenAI said it is planning its next internal security sprint with Cathedral Systems, whose researchers will use OpenAI's models alongside its own teams to challenge its defenses. The same day, it raised critical-finding bounties in ChatGPT, Codex and the API Platform to as much as $30,000, with exceptional reports eligible for up to $172,000 including bonuses. The sprint builds on an earlier one that assigned 25% of product engineers to security work. OpenAI has not said when the Cathedral work starts, how long it runs, which systems are in scope or whether Cathedral is paid.
Cathedral was founded in 2026 by four former Department of Government Efficiency staffers. A draft report reviewed by The Washington Post says its AI agents reached Anthropic's internal systems in under 12 hours by exploiting third-party business software during a bug bounty test. Anthropic confirmed access to a limited set of sales, HR and IT systems, but says the work took several days, that Cathedral broke program rules by not disclosing all its activity, and that model weights and Claude conversations were not affected.
On November 3, Reuters reported, citing five people familiar with the matter, that employees had warned OpenAI's leaders and board about weak internal security for years, beginning before ChatGPT launched in late 2022. The warnings covered broad code access and internal tools, such as model-testing sandboxes, that were not regularly security checked. OpenAI disputed that it had neglected security, saying its most sensitive work sits in separate repositories with strict access controls.
Why it matters
For founders, model providers now use their own models as offensive testers, so outside red-team capacity is becoming a standing cost of shipping AI products. Cathedral has reportedly raised $160 million at a $1.4 billion valuation from Andreessen Horowitz and Sequoia Capital, which shows investors are funding firms built around this kind of testing. The $30,000 and $172,000 bounty ceilings also set a visible price for what independent researchers can earn on AI products.
For builders, the Anthropic episode points to a practical audit. Map who can reach model weights, evaluation data, internal tooling and vendor-connected business systems, because the reported entry point was third-party software rather than the model stack itself. Treat outside red-team access as a standing process with written scope, disclosure rules, logging and a way to cut access quickly. Investors should ask portfolio companies the same questions about access reviews and vendor inventories.
Key points
- 1OpenAI plans an internal security sprint in which Cathedral Systems researchers use OpenAI's models against its defenses.
- 2Critical bugs in ChatGPT, Codex and the API Platform can now earn up to $30,000.
- 3Exceptional findings can earn up to $172,000 including bonuses, and the new rates apply to pending reports.
- 4A draft report says Cathedral reached Anthropic's sales, HR and IT systems in under 12 hours via third-party software.
- 5
