OpenAI raises the top critical bug payout to $172,000

OpenAI is preparing another internal security sprint with researchers from Cathedral. They will work alongside internal teams and use OpenAI models to find weaknesses, test fixes and improve continuous defense.

Cathedral will turn OpenAI models against OpenAI systems

The company is building on a previous sprint that assigned 25% of product engineers to finding and fixing weaknesses. Codex helped investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for review.

The Security Bug Bounty is changing at the same time. A critical vulnerability in ChatGPT, Codex or the API Platform can now earn up to $30,000. An exceptional critical finding retains a $100,000 base ceiling, but bonuses can lift the payout to $171,925, rounded to $172,000.

Bigger payouts buy better reports and longer researcher relationships

OpenAI is adding bonuses for reproducible reports, a 5 tier loyalty program and temporary campaigns for new features. The money is meant to reward more than discovery. It encourages researchers to provide usable evidence, explain impact and return for further testing.

For teams deploying AI agents, the operating model matters most. A company model may accelerate defense, but it still needs isolated environments, human review and outside people willing to challenge internal assumptions.

A company model playing attacker is not an independent audit

The announcement does not say which systems Cathedral will test, how long the sprint will run or how many vulnerabilities the previous sprint found. OpenAI also selects the tools, environment and scope. An outside partner adds another perspective without making the whole process independent.

A larger bounty measures incentives rather than security. The $172,000 figure requires a combination of the base award and several bonuses in an illustrative scenario. An ordinary critical flaw in a core product has a $30,000 ceiling.

Published fixes and findings outside the sprint will settle the case

Continuous defense will become measurable if OpenAI publishes the classes of bugs found, remediation speed and results from repeated testing. High quality reports from researchers outside the paid sprint will provide a second signal.

The program becomes convincing when models help close weaknesses before a hostile actor reaches them and the fixes survive another round of testing.

Lilith's verdict

OpenAI is sending its own models to guard its own vault while Cathedral tries to take the keys. Trust arrives only when the company shows the cracks it found and fixes that survived another blow.