OpenAI Plans Its Next Internal Security Sprint With Cathedral, More Than Quadruples Critical Bug Bounty Rewards to Up to $172,000

Background and the problem

AI is getting better at finding vulnerabilities, and both attackers and defenders are using it. OpenAI argues that traditional cyber defenses alone are no longer enough and that defenders need frontier models first, running a continuous loop of finding and fixing weaknesses. In late October, Cathedral ran a widely discussed security test against Anthropic; OpenAI is now bringing the same firm in on purpose, a clear sign that leading labs are openly inviting outside teams to attack them with AI.

How the program works

The sprint has three parts. First, Cathedral researchers will work with OpenAI teams, using OpenAI models to challenge its systems; Cathedral CEO Gavin Kliger says the goal is to find weaknesses and test fixes. Second, the Security Bug Bounty program changes, with rewards for critical findings more than quadrupled to encourage outside researchers to challenge OpenAI's assumptions. Third, it builds on the last sprint, which per OpenAI dedicated 25% of product engineers and used Codex to investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for human review. OpenAI also offers its most capable models to Daybreak customers and has published the architecture and lessons from the last sprint.

Benchmarks

This is a security program announcement, so there are no model benchmarks. OpenAI's figures: critical rewards up more than fourfold, up to $172,000 including bonuses per its X post, 25% of product engineers on the last sprint, and $1 billion in subsidized access over six months. How many issues the last sprint fixed, and the timing and scope of the next one, were not disclosed.

Getting started

Security researchers should review the updated OpenAI Security Bug Bounty rules, since critical findings now pay much more; check scope and disclosure terms before submitting. Enterprise security teams can borrow the approach: run a focused sprint with a slice of engineers, let coding agents reproduce bugs and draft patches in isolation, and keep humans reviewing every merge. Organizations protecting essential infrastructure can look into the $1 billion subsidized access program.