OpenAI plans new Security Sprint and increases Bug Bounty rewards – Unite.AI

OpenAI announced on November 2, 2026 that it is planning another internal security sprint with Cathedral and other external researchers, while a bug-bounty update on the same day garnered rewards and introduced bonuses that can bring bumper critical payouts to $172,000.

Cathedral researchers will use OpenAI’s models alongside its internal teams to challenge defenses and look for weaknesses. The announcement is for a planned sprint, not a completed test cycle. Gavin Kliger, CEO and co-founder of Cathedral, said the collaboration will pair OpenAI’s security models and expertise with Cathedral’s research to identify weaknesses, test fixes and support ongoing defense.

Bigger rewards for independent researchers

According to OpenAI’s separate bug bounty announcement, the changes more than quadrupled the midpoint of the reward ranges for medium, high, and critical results. A critical vulnerability in a major product such as ChatGPT, Codex or the API platform can now fetch up to $30,000. Exceptional critical achievements on such objectives remain eligible for up to $100,000 before bonuses.

High quality reports will be entitled to a multiplier on the base size. The quality bonus recognizes a robust proof of concept that isolates the bug, shows when it can be exploited, and demonstrates its impact. A new five-tier loyalty program adds scalable bonuses for researchers who repeatedly work with OpenAI, while priority campaigns will provide dedicated bonus pools for testing newly launched features.

Reports still awaiting a premium will use the new ranges. Researchers whose rewards were awarded in the 30 days before the announcement will receive the difference when the new base and loyalty rewards would have paid more.

The previous sprint informs the defense factory

The previous sprint provides the operational model. OpenAI dedicated 25% of its product engineers to finding and fixing weaknesses, using Codex to investigate findings, reproduce vulnerabilities in isolated environments, and prepare patches for review. Its Defense Factory account says the initiative has mobilized more than 250 people in more than 100 service areas.

OpenAI reports that the previous effort solved 53 urgent or high-priority issues in the first day. It identified 37% of results as duplicates, reproduced 19.5% of them at runtime, and had a false positive rate of 0.81% after dynamic validation. The repair was entirely Codex based, with an override rate of 0.53%.

These results explain why Defense Factory is more than a patch generator. Its continuous, agent-centric loop covers inventory, discovery, dynamic validation, ownership assignment, and verified remediation, with agents working in reproducible, isolated environments. People review consequential changes and handle ambiguous cases while officers undertake routine work.

The project also involves a material qualification: follow-up checks highlighted a gap between the merged patches and the fixes implemented across the fleet. Moving the ticket or accepting work does not prove repair; the verified remediation phase requires implementation and verification testing after human review and authorized implementation.

Defender access beyond OpenAI

In addition to its own systems, OpenAI is making its most capable models available to Daybreak customers for authorized defensive work. Daybreak’s September 3, 2026 announcement set a six-month consumption goal for a $1 billion commitment covering subsidized access, training, technical support and partnerships, starting with U.S. defenders of resource-constrained essential services like water and electricity.

At the time of that announcement, OpenAI reported that thousands of defenders across 2,000 approved organizations and workspaces were already using Daybreak, including cybersecurity companies, defense organizations, and law enforcement agencies.

The planned sprint bridges those internal and external defenses: External researchers receive more incentive to report what OpenAI’s tests missed, while its agent-first operation is built to validate and fix weaknesses. The benchmark built into OpenAI’s design is whether the results end in an independently implemented and verified fix rather than tickets simply being marked as complete.