OpenAI plans Cathedral sprint, bug bounties up to $172,000

OpenAI on November 2 raised the rewards in its Security Bug Bounty program, more than quadrupling the midpoint of its medium, high and critical reward ranges on average. In a post on the OpenAI blog, it also said researchers from Cathedral will join its next internal security sprint. If you have a report that hasn’t been awarded yet, it will be paid under the new ranges. If your bounty was awarded in the 30 days before the announcement, OpenAI will pay the difference when the finding would have earned more.

That top-up covers the base award plus loyalty awards, according to the bounty announcement. Bounties older than 30 days fall outside it.

What the new ranges pay

A critical vulnerability in a core product, meaning ChatGPT, Codex or the API Platform, now earns up to $30,000. Exceptional critical findings on those targets are still capped at a $100,000 base award. OpenAI set that ceiling in an earlier round, when it lifted the maximum payout from $20,000. The program has run on Bugcrowd since April 2023. So the new headline figure of up to $172,000 comes from bonuses on top of an unchanged base. OpenAI’s own example reaches $171,925 for an exceptional finding, such as being able to read arbitrary ChatGPT users’ chat messages. To get there it stacks a quality bonus, a priority campaign bonus and the top loyalty tier, which is named 31337.

The worked examples show the old and new payouts side by side. A one-click account takeover in ChatGPT, which historically paid $6,500, comes out at $51,577 with all three bonuses applied, almost eight times as much. A medium bug, where a read-only member could upload files to a ChatGPT Project without write permission, goes from $750 to $2,415. OpenAI calls all of these scenarios and historical awards illustrative.

How the bonuses stack

The quality multiplier pays for reports that give OpenAI a head start on a fix. That means a reliable proof of concept that isolates the bug, shows when it can be exploited and demonstrates its impact. In every example it adds 15% to the base bounty.

The loyalty program has five tiers and rewards researchers who keep reporting. The examples show Bronze at 1.05, Gold at 1.10 and 31337 at 1.15. Priority campaigns are short-term bonus pools for newly launched features, and they carry a 1.30 multiplier in the examples.

The sprint with Cathedral

Cathedral’s researchers, named Cathedral Systems in OpenAI’s post on X, will use OpenAI’s models alongside its own teams to challenge its defenses. In the last sprint, OpenAI put 25% of its product engineers on finding and fixing weaknesses. Using Codex, they investigated findings, reproduced vulnerabilities in isolated environments and prepared patches for review. In September OpenAI described that loop as its Defense Factory: agents find vulnerabilities, validate them and check that the fixes work, with more than 250 people working across hundreds of its systems at the time. The bounty post presents outside reports as a complement to that work, aimed at vulnerabilities its own testing may have missed.

Beyond its own systems, OpenAI says it offers its most capable models to Daybreak customers. It has also committed $1 billion in subsidized access over six months for frontline defenders protecting essential services and digital infrastructure. It is publishing the architecture and lessons from the last sprint, along with a blueprint for cyber defenders in industry and government.

Researchers can test targets in person on November 20-21 at PROMPT2PWN in Oakland, where OpenAI’s security team will attend.