OpenAI raises its top reward for an exceptional critical flaw to $172,000

OpenAI is preparing another internal security sprint with researchers from Cathedral Systems. They will use OpenAI models alongside company teams to find weaknesses, reproduce vulnerabilities and prepare fixes. Rewards for critical findings are increasing by more than four times, with exceptional reports eligible for up to $172,000 including bonuses.

External researchers will turn OpenAI models against OpenAI systems

Cathedral will bring its own methods and challenge the company's defenses alongside internal teams. The announcement builds on a previous sprint that involved 25% of OpenAI product engineers. Codex helped investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for review.

The company is also making its strongest cyber capabilities available to Daybreak customers. It has committed $1 billion in subsidized access over six months for defenders of essential services and digital infrastructure.

Security is moving from periodic audits to a continuous defensive loop

OpenAI's argument rests on shortening the time between discovery and repair. A model can inspect more code and assist reproduction, but a human still needs to assess impact and approve a patch. An external team can also challenge blind spots shared by the company's own developers.

The higher bounty matters as much as model capability. When legitimate researchers receive a meaningful reward and a clear disclosure channel, the company improves the odds that critical information reaches defenders first.

The announcement gives no start date or measurable definition of success

OpenAI has not disclosed the sprint's timing, duration, system scope or target repair time. It also does not say how many vulnerabilities the previous sprint found or how many would have escaped without models. That makes it hard to separate better security from an impressive mobilization of people and compute.

Models can accelerate both sides. Access controls for trusted defenders and isolated environments reduce misuse risk, but they do not eliminate leaked techniques or false findings that overload triage.

Time from disclosure to patch will reveal the sprint's value

The useful metrics are confirmed critical findings, false positive rates, median repair time and regressions introduced by patches. It will also matter whether OpenAI publishes methods and lessons that defenders can use beyond its own platform.

A $172,000 reward attracts attention. Trust will come from predictable report handling, rapid fixes and clear rules for researchers testing the company's defenses in good faith.

Lilith's verdict

OpenAI is inviting researchers to aim its own models at its systems and putting $172,000 on the table. It now needs to show that the repair crew arrives before the press release.