OpenAI plans another internal security sprint, this time bringing in outside researchers from Cathedral Systems to test the company’s own systems with the help of its AI models. Announced on November 2, the effort arrives alongside a much larger bug-bounty program. It is a notable shift from treating AI mainly as a product challenge to using it as part of the security testing loop—but the sprint is still a plan, not a published result.
What OpenAI announced
In its announcement, OpenAI said Cathedral researchers will work alongside its teams and use OpenAI models to challenge defenses and look for weaknesses. Cathedral co-founder Gavin Kliger described the collaboration as combining the company’s models with outside security expertise. OpenAI has not published a schedule, scope, or a list of systems that the new sprint will test, so those details remain unknown.
A bigger incentive for responsible disclosure
OpenAI also raised rewards in its Security Bug Bounty program. The company says average midpoint rewards for medium, high, and critical findings have increased by more than four times. A critical vulnerability affecting a core product such as ChatGPT, Codex, or the API Platform can now receive up to $30,000. Exceptional critical reports can qualify for as much as $100,000 before bonuses—and up to $172,000 when eligible bonuses are applied.
The new structure adds quality, loyalty, and priority-campaign multipliers. OpenAI’s published example calculates a $100,000 exceptional award at $171,925 after applying the stated multipliers; the headline “up to $172,000” is a rounded maximum, not a standard payment. Reports awaiting an award are covered by the new ranges, and OpenAI says qualifying awards from the preceding 30 days will receive the difference where applicable. Researchers still need to follow the program’s scope and disclosure rules.
Why an internal sprint is different from a bug bounty
A bounty program invites independent researchers to report flaws under defined rules. An internal sprint is a coordinated exercise: teams can select systems, reproduce findings in controlled environments, and route fixes through engineering review. OpenAI says its previous sprint devoted 25% of product engineers to finding and fixing weaknesses, with Codex used to investigate findings and prepare patches for review. That is the company’s account of the earlier effort, not independent evidence that every weakness was found or fixed.
AI can speed testing—and widen the consequences
AI-assisted analysis can help security teams search code, generate test cases, and investigate a suspected weakness more quickly. The same capabilities can also help attackers understand unfamiliar systems or adapt known techniques. That makes the boundary between useful testing and unsafe access especially important: models need tightly scoped permissions, isolated test environments, logging, and human review before a proposed change reaches production.
The sprint does not mean a model is being given unrestricted access to OpenAI’s live infrastructure. The company’s announcement describes a defensive exercise but does not publish the exact controls, environments, or access boundaries. Organizations adopting similar workflows should treat those as design requirements, not assumptions.
The operational test is what happens after a finding
Finding a vulnerability is only the beginning. Security value comes from confirming that a report is reproducible, assessing its real impact, assigning an owner, shipping a safe fix, and checking that the issue stays closed. OpenAI says its earlier sprint used isolated environments to reproduce vulnerabilities and prepare patches for review. For the new round, the public announcement does not yet provide metrics such as confirmed issues, remediation time, or recurrence rates.
What readers should watch next
The meaningful follow-up will be evidence: a clear account of the sprint’s scope, how external findings were validated, what categories of weaknesses were fixed, and whether the bug-bounty changes attract useful reports. OpenAI’s November 2 announcements set out new commitments and incentives, but do not disclose results from this planned sprint.
For technology companies building increasingly capable AI, security cannot be a final checkpoint after development. OpenAI’s approach puts continuous testing and outside scrutiny closer to the development process. Whether it improves protection will depend less on the announcement than on the controls, fixes, and transparent follow-through that come next.
