OpenAI Plans AI-Assisted Security Sprint, Raises Bug Bounties

OpenAI plans a new security sprint

OpenAI says it is planning a new internal security sprint that will bring outside researchers together with company teams to test and strengthen its systems. The company is partnering with Cathedral Systems, whose researchers will use OpenAI models alongside OpenAI’s own teams to look for weaknesses and help test defenses.

The announcement signals a combined approach: coordinated security research, internal engineering work and AI-assisted testing. OpenAI has not disclosed a schedule, the systems or products that will be in scope, or any results from the upcoming sprint. It should therefore be understood as a planned exercise, not a completed test or a report of vulnerabilities already found.

The announcement came alongside a substantial change to the company’s Security Bug Bounty program. OpenAI says it is increasing rewards for medium, high and critical findings, with the midpoint of reward ranges across those categories rising by more than four times on average. The company says the changes are intended to encourage researchers to challenge its assumptions and help improve its defenses.

Cathedral researchers will work with OpenAI teams

Cathedral’s researchers are expected to work with OpenAI teams, using the company’s models to challenge its defenses and help assess fixes. The partnership adds an external research perspective to an internal security effort, though OpenAI has not described the specific testing methods, targets or safeguards that will apply to the new sprint.

That lack of detail matters. The announcement does not identify a start date or duration, nor does it say whether the work will focus on particular products, infrastructure or classes of vulnerability. OpenAI has also not published a scope document or outcomes for the planned exercise. Until those details are available, researchers should rely on the published bug bounty program rules for authorized testing rather than assuming the sprint creates a broader testing permission.

The company framed the partnership as part of a wider effort to use AI to find and fix weaknesses. Its statement argues that traditional cyber defenses alone are no longer sufficient, emphasizing the need for organizations to use advanced capabilities to strengthen their systems. For software teams, the practical question is not simply whether a model can suggest a test, but whether a finding can be safely reproduced, assessed and turned into a reliable fix.

Previous sprint involved a quarter of product engineers

OpenAI pointed to its last internal security sprint as evidence for the approach. The company said it dedicated 25% of its product engineers to finding and fixing weaknesses in its systems. Engineers used Codex to investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for review.

That description offers a limited view of how AI was used in the earlier effort: as assistance for investigation and reproduction, followed by patch preparation for human review. OpenAI has not provided a full accounting in the new announcement of vulnerabilities discovered, how many fixes were deployed or the time required to resolve issues. The upcoming sprint’s results remain unknown.

Reproducing a potential vulnerability in an isolated environment can help determine whether a report is actionable without exposing production systems to unnecessary risk. Preparing a patch for review can also shorten the path from diagnosis to remediation. But a proposed change still needs engineering review and validation in the context of the affected service; the announcement does not suggest that a model-generated patch is automatically deployed.

For builders, that workflow is familiar even outside AI security: detection is only the first step. A useful security process must connect discovery with triage, safe reproduction, ownership, remediation and testing. OpenAI’s description of its previous sprint highlights some of those stages, while leaving the final results undisclosed.

Critical findings in core products can earn up to $30,000

Under the updated rewards, a critical vulnerability affecting a core product such as ChatGPT, Codex or the API Platform can earn up to $30,000. OpenAI says exceptional critical findings on those targets remain eligible for a base award of up to $100,000.

The company’s example shows how bonuses can raise the total. An exceptional ChatGPT finding that significantly compromises a security boundary—such as enabling access to other users’ chat messages—has a listed base amount of $100,000. Applying the example’s quality, priority-campaign and loyalty multipliers produces a total of $171,925.

That figure is an exceptional case, not the standard payout for a critical report. Actual awards depend on the finding and applicable program criteria. OpenAI describes the broader change as more than quadrupling the midpoint of reward ranges for medium, high and critical reports on average; that does not mean every submission will receive four times its previous award.

Bonuses and retroactive adjustments widen the program

The reward structure combines a base award with possible additional incentives. The published example applies a 1.15 quality multiplier, a 1.30 priority-campaign multiplier and a 1.15 loyalty multiplier to the $100,000 base award. In that example, the combined result is $171,925.

OpenAI has also said that researchers who submitted reports during the 30 days before the change can receive the difference between the old and new rates retroactively. That policy may affect eligible recent submissions, but it does not mean every report from that period automatically qualifies for the maximum amount. Researchers will need to refer to the program’s criteria and the company’s assessment of each finding.

The company says the revised incentives are intended to attract useful security research and reward work that helps protect its products. For researchers, the practical value of a bounty program also depends on clear scope, safe-harbor terms, predictable triage and timely communication—details that determine how a vulnerability can be tested and reported responsibly. The size of a headline reward does not replace the need to understand the program’s rules before testing.

What remains unknown about the upcoming sprint

OpenAI has announced a partner and an intended purpose, but has not published the operational details needed to assess the sprint’s coverage. There is no announced timeline, defined scope or account of results for the planned work. The company’s description of the previous sprint provides a process outline, not a substitute for findings from the next one.

The bounty changes and the sprint are related parts of a broader security effort, but they serve different roles. The sprint brings outside researchers and internal teams together; the bug bounty program offers a channel for researchers to report qualifying vulnerabilities under its rules. OpenAI has not said that participation in the bounty program is a prerequisite for the sprint or that the new collaboration changes the program’s terms.

For developers using OpenAI services, the announcement does not change API behavior, product security settings or integration requirements. Its significance is about how the provider says it intends to test and improve its own systems, and how it is trying to attract external vulnerability reports. More meaningful assessment will be possible when the company shares the sprint’s scope or results.

**Source date: November 2, 2026.**