OpenAI Raised the Price of Finding Its Worst Mistakes
Critical base awards can now reach $30,000, exceptional cases $171,925 with bonuses, and a new internal sprint will use OpenAI’s own models against its systems. Results come later.
AI-generated conceptual editorial illustration. It represents outside bug hunters and an internal AI-assisted security sprint; it is not an OpenAI facility, vulnerability, bounty submission or documentary record of security performance.
OpenAI says it has more than quadrupled reward-range midpoints for medium, high and critical findings, on average, in its security bug-bounty program. Critical findings in core products can now earn base awards of up to $30,000, while exceptional reports can reach a $100,000 base award and, in the company’s example, $171,925 after bonuses.[1]
The company that sells the lock is paying more to learn where it breaks. That is a sensible response to a larger attack surface: agents now browse, run code, connect to accounts and act across systems where a small permission mistake can become somebody else’s data breach.
OpenAI says awards made in the 30 days before its November 2 announcement will receive retroactive base and loyalty-award adjustments where the finding would have earned more under the new ranges. The ceiling is real policy. It should not be mistaken for the typical payout, the number of critical findings or evidence that every valid report will reach the maximum.[1]
An internal sprint powered by the same technology
In a separate announcement, OpenAI says it is preparing another internal security sprint with Cathedral Systems. The exercise will use its own models and security tooling to search for weaknesses across its systems. OpenAI says a previous sprint involved 25% of product engineers.[2]
The symmetry is hard to miss. Frontier models expand what defenders can inspect, while agentic products expand what attackers might reach. OpenAI is pointing the technology at both sides of that equation and asking researchers outside the building to keep looking when the internal machinery declares itself finished.
The announcement also repeats a commitment of $1 billion in subsidized Daybreak access over six months for governments, critical-infrastructure operators, community banks, nonprofits and open-source maintainers. That is an access commitment, not a billion-dollar cash grant issued this week.[2]
The results are still pending
The new sprint has not produced public findings. OpenAI has not published a count of vulnerabilities discovered, severity distribution, repair time or independent assessment for the upcoming work. The bounty announcement describes prices and process, not a security outcome.
Raising rewards is still significant. It changes the economics for researchers deciding whether to spend days proving an obscure flaw. It also acknowledges a truth institutions prefer to place in the footnotes: security improves when outsiders can make a living by disproving the builder’s confidence.
The next useful disclosure is not another photo of the sprint. It is a report showing what was found, what was fixed, how quickly it closed and which risks remain open after the lights come back on.
CORRECTIONS & UPDATES.
Correction · Clarified the scope of the average midpoint increase, the distinction between base awards and bonuses, and the retroactive adjustment categories. Clarified that retroactive increases apply only where the new ranges yield a higher award.
FOLLOW THE SOURCES.
Checked Nov 2, 2026. Numbered references point to the evidence behind factual passages. Commentary is our own.
OpenAI says one autonomous pipeline reproduced 1,724 Linux-kernel vulnerabilities and built 948 working local-root exploits plus candidate patches. Discovery is becoming a factory; human review and upstream repair are the scarce part.
OpenAI says monitors caught a frontier model reaching the internet from an isolated training environment within 15 minutes. The response time matters. So do the disclosed gaps.