The OpenAI bug bounty programme will now pay external security researchers up to $172,000 for exceptional vulnerabilities, as the company prepares an internal security sprint with outside firm Cathedral Systems. The payout updates more than quadruple the average reward ranges for medium, high and critical security flaws across its main artificial intelligence systems.
Key points
- OpenAI has raised baseline critical bug bounties to $30,000, with exceptional findings eligible for up to $172,000 after multipliers and bonuses.
- Outside firm Cathedral Systems will join internal engineers during an upcoming security sprint to stress-test systems and test fixes.
- The move builds on the company’s automated Defense Factory, which used Codex to resolve dozens of urgent flaws during an earlier internal test.
- Researchers who received rewards in the 30 days before the announcement will receive retroactive pay increases under the new rates.
OpenAI revealed the changes on 2 November 2026, combining higher financial rewards for independent bug hunters with a broader push to incorporate third-party researchers into internal engineering workflows. The updated incentives target flaws across flagship products, including ChatGPT, Codex and the developer API Platform.
Under the revised fee schedule described in OpenAI’s separate bug-bounty announcement, a standard critical vulnerability in a primary product pays up to $30,000. Exceptional critical submissions can earn a base reward of up to $100,000. New multipliers for detailed proof-of-concept reports, paired with loyalty bonuses and dedicated feature campaigns, lift the top theoretical payout to roughly $172,000.
How the revised bounty structure works
The updated programme introduces several bonus mechanisms aimed at keeping talented researchers focused on OpenAI’s services. Standard reward ranges for medium, high and critical submissions have increased more than fourfold at their midpoint values.
Reports that supply an isolated proof of concept detailing practical exploitability receive a quality multiplier on the baseline payout. OpenAI also established a five-tier loyalty scheme that rewards researchers who consistently file verified reports over time. Temporary priority campaigns will also offer dedicated cash pools to scrutinise newly launched products.
Researchers with reports currently under review will be evaluated under the new rates. OpenAI said it will also pay the difference retroactively to researchers who received bounties during the 30 days preceding the announcement, matching what the new baseline and loyalty tiers would have delivered.
Cathedral partnership and the planned sprint
Alongside higher payments, OpenAI outlined an upcoming security sprint in a published defence update. External researchers from cybersecurity firm Cathedral Systems will work alongside OpenAI engineers to identify vulnerabilities and test system remediations.
OpenAI has not published a start date for the sprint. The company confirmed that Cathedral’s analysts will be given access to internal frontier models to evaluate software protections and probe for logical flaws across internal architecture.
Gavin Kliger, chief executive and co-founder of Cathedral, said the collaboration would pair OpenAI’s models and security expertise with his team’s research to “find weaknesses, test fixes and support continuous defense”.
The joint exercise is designed to challenge developer assumptions before product rollouts. OpenAI noted that partnering with external firms provides adversarial perspectives that purely internal product teams often miss during routine development cycles.
Lessons from the previous internal sprint
The planned collaboration follows an extensive internal exercise where OpenAI assigned 25 percent of its product engineers to vulnerability discovery and patch remediation. That initiative formed the foundation of what OpenAI calls its Defense Factory.
According to OpenAI’s Defense Factory account, the prior sprint mobilised more than 250 engineers across more than 100 service divisions. The internal team resolved 53 urgent or high-priority vulnerabilities within the first 24 hours of the operation.
Automation played a dominant role during that internal assessment. Engineers used Codex to investigate bug reports, replicate runtime vulnerabilities in isolated sandboxes and draft pull requests for human review. The company reported that 19.5 percent of findings were reproduced at runtime, while 37 percent were identified as duplicate reports.
Dynamic validation produced a false-positive rate of 0.81 percent. All software remediation across that exercise relied on Codex-generated patches, which recorded a rollback rate of just 0.53 percent.
The company acknowledged operational hurdles during the process. Follow-up checks revealed a distinct gap between merging a software patch and actually deploying that fix across live production clusters. As a result, the current Defense Factory architecture mandates post-deployment verification evidence before any vulnerability ticket is formally marked as resolved.
Expanding AI tools for frontline defenders
OpenAI’s focus on automated system defence mirrors its wider operational strategy, which includes public-sector alignment such as when OpenAI hires Dan Driscoll to help steer national security modernisation projects.
The company has argued that manual vulnerability tracking cannot keep up with current cyber threats. To offset that imbalance, OpenAI has begun supplying defensive tools to critical infrastructure operators outside its own corporate boundaries.
Under its September 3, 2026 Daybreak announcement, OpenAI committed $1 billion over six months in subsidised access, training and technical resources for frontline defenders. The project gives priority to under-resourced United States organisations managing critical civil infrastructure, including power grids and municipal water systems.
More than 2,000 verified organisations and workspaces have received Daybreak access, including commercial cybersecurity providers, intelligence units and law-enforcement bodies. These external partners use frontier AI models to parse system telemetry, identify anomalies and isolate malicious code in defensive environments.
The competitive landscape for vulnerability rewards
The revised bounty tiers bring OpenAI’s top compensation closer to traditional technology giants, though differences remain based on exploit scope. A $172,000 ceiling places OpenAI near Google’s general vulnerability reward brackets for web services and cloud infrastructure.
However, major platform vendors like Apple and Microsoft still maintain larger top-line bounties, which can exceed $1 million for full-chain, zero-click kernel compromises and operating system exploits. OpenAI’s programme remains tailored specifically toward application logic, API abuse, data leakage and jailbreak-driven exploit chains.
By pairing higher monetary bounties with direct model access for partners like Cathedral, OpenAI hopes to locate systemic software flaws before adversaries can target them in production. The company confirmed it will publish technical takeaways once the joint sprint concludes.
