OpenAI raises bug bounties, with exceptional awards near $172,000
OpenAI raised its security bug bounty rewards on November 2. Critical flaws in core products including ChatGPT, Codex and the API Platform can earn up to $30,000 before bonuses. Exceptional findings retain a $100,000 base ceiling, with an illustrative maximum of $171,925 after bonuses.
What the larger payout means
The roughly $172,000 headline combines quality, campaign and loyalty incentives. It is not the standard critical award. OpenAI explicitly labels its scenarios and historical awards illustrative, so they should not be read as disclosures of actual attacks. Read OpenAI’s announcement.
Unawarded reports receive the new ranges. Eligible bounties awarded during the preceding 30 days receive the difference under the new base and loyalty awards.
Which reports qualify
The current Bugcrowd policy excludes jailbreaks, hallucinations and safety bypasses limited to model behavior. A model-assisted attack can qualify when it exploits a separate flaw in an eligible OpenAI product or controlled integration. Researchers need to identify the control that failed and demonstrate the unauthorized result using their own test data.
A striking chat response and a failed application permission check are different claims. Evidence should establish which one occurred.
Bonuses depend on the submission
| Mechanism | Maximum | Condition |
|---|---|---|
| Report quality | 1.15× | Clear proof, prerequisites and impact |
| Priority campaign | 1.30× | Qualifying designated campaign |
| Loyalty | 15 percent extra | Highest loyalty tier |
For the exceptional example, $100,000 × 1.15 × 1.30 × 1.15 = $171,925.
The factors multiply rather than simply add together. A difficult-to-understand report can receive a downward adjustment. Final ratings and rewards remain at OpenAI’s discretion.
Testing must stay within the listed scope. Researchers should use their own accounts and stop if another person’s data appears, rather than collecting more records as proof.
November 6 update: Hacktron targets service connections
OpenAI added a research partnership on November 6, announcing that its engineers and Hacktron researchers will jointly choose high priority areas to test. The focus is how OpenAI’s systems interact with third party services. OpenAI is also expanding Hacktron’s access to frontier models through Daybreak Red. Both OpenAI’s announcement and Hacktron’s statement describe the collaboration.
For bounty participants, the important distinction is access. This is a coordinated engagement with Hacktron; the announcement does not grant every researcher the same model access. OpenAI’s Daybreak overview describes Red as access to specialized cybersecurity models for authorized research, exploit validation and security testing, with approval and use restrictions.
The partnership follows the November 2 security sprint announcement, which named Cathedral as a participant alongside OpenAI’s teams. These are distinct efforts: the public bounty program accepts eligible reports, the planned sprint brings outside researchers into an internal exercise, and Hacktron will help prioritize testing of service connections. Neither partnership announcement provides results from the newly announced work.
Service connections deserve attention because an agent can move between applications during one task. For an independent researcher, that makes authorization boundaries especially important. A collaboration focused on integrations gives no blanket permission to test every connected provider. Check the permitted target and actions before beginning, and keep the report tied to reproducible evidence.
Our separate guide to OpenAI’s Linux security study and patch review explains why research findings, accepted bounty submissions and fixes delivered to users need different evidence.
The practical takeaway
ByteForward’s view is that reproducible evidence matters more than the headline ceiling. Read the applicable target rules before planning research, then estimate compensation only for conditions the report actually meets.
