OpenAI is increasing its security bug bounty rewards, with exceptional critical vulnerabilities in core products potentially earning researchers approximately $172,000 after qualifying bonuses.
The company announced the higher payouts on Nov. 2 alongside plans for another AI-assisted security sprint, bringing Cathedral researchers together with internal teams to find and fix vulnerabilities.
Results from the previous exercise highlight a challenge for security teams: faster vulnerability discovery still requires reliable validation, clear ownership, and verification that fixes reach production.
Earlier security testing closed 53 urgent or high-priority issues
Researchers from Cathedral will join internal teams to examine defenses and evaluate fixes. Participants will use OpenAI’s models with their own research expertise to investigate potential weaknesses.
During the previous sprint, more than 250 people worked across over 100 service areas, with 25% of product engineers dedicated to finding and fixing vulnerabilities. OpenAI said teams closed 53 urgent or high-priority issues on the first day.
Codex helped engineers investigate suspected flaws and reproduce them in isolated environments before preparing patches for human review. Engineers also developed processes for assessing severity and routing findings to responsible teams, extending traditional vulnerability testing with automated investigation.
OpenAI said its triage process identified 37% of findings as duplicates, while runtime validation reproduced 19.5%. Findings that did not reproduce were not necessarily false positives; the company also described testing constraints involving missing dependencies and configuration differences. Engineers used those experiences to develop the Defense Factory, a continuous process connecting vulnerability discovery with validation and verified remediation.
Higher bounties reward serious and well-documented vulnerabilities
OpenAI is increasing payouts under its Security Bug Bounty program. Midpoint rewards for medium-, high-, and critical-severity findings are rising more than fourfold on average.
Critical flaws in ChatGPT and other core products can now earn up to $30,000 under standard reward ranges. Exceptional critical vulnerabilities in those core products remain eligible for base awards of up to $100,000, with qualifying bonuses increasing potential payouts to approximately $172,000.
Researchers participating in bug bounty programs can also qualify for additional incentives under the revised structure.
- Report-quality bonuses. Higher payouts reward detailed submissions with reliable proof of concept, including evidence demonstrating when a vulnerability can be exploited and its potential impact.
- Researcher loyalty rewards. A five-tier program offers increasing bonuses to researchers who consistently contribute security findings.
- Priority campaign incentives. Temporary campaigns provide additional rewards for investigating newly released features.
Pending submissions will qualify for the updated reward ranges. For bounties awarded during the 30 days before the Nov. 2 announcement, OpenAI will pay the difference if the finding would have earned more under the updated base-award and loyalty-award structure.
Faster vulnerability discovery puts pressure on remediation
OpenAI's earlier sprint exposed a challenge for organizations adopting AI-assisted vulnerability scanning.
Automated detection can produce findings faster than security teams can investigate them, particularly when duplicate reports and unconfirmed weaknesses compete for attention. In large enterprise environments, time spent sorting inconclusive reports can delay work on verified vulnerabilities affecting exposed systems.
OpenAI said follow-up checks exposed a gap between merged patches and fixes deployed across its fleet, prompting it to expand verification. That experience illustrates why confirming a vulnerability and preparing a patch are only part of remediation. Security engineers may identify a vulnerability, but application teams still need to test and deploy the correction. A patch approved by developers may leave organizations exposed if older versions continue running on production servers or containers, creating a gap between completed development work and actual remediation.
For security leaders evaluating AI-assisted testing, a more useful measure is how long confirmed vulnerabilities remain exploitable in production. Delays during investigation can reveal problems validating findings, and fixes awaiting deployment expose weaknesses in release management. Tracking those delays through patch verification can help organizations determine whether faster discovery is reducing the time their systems remain vulnerable.
Related reading: Past cyberattacks can reveal security weaknesses, but AI-powered threats may demand a different approach. Learn how organizations can prepare.