OpenAI Safety Bug Bounty on Bugcrowd
OpenAI pays $250 to $7,500 per accepted report for safety and abuse flaws in its agentic products and platform. Public on Bugcrowd, open all year.
The OpenAI Safety Bug Bounty is a public program on Bugcrowd. OpenAI launched it on 25 March 2026. It pays for safety and abuse issues in OpenAI products, even when they do not count as a normal security vulnerability. Rewards depend on severity. The top level pays $5,500 to $7,500 and the lowest paid level pays $250 to $500. OpenAI decides the final reward.
There are four areas in scope. The first is agentic tools, including MCP and connectors. This covers products that act for the user or read data as the user, such as Atlas Browser, Codex, Operator and agentic ChatGPT tools. Prompt injection that hijacks an agent must work at least half of the time. The second is OpenAI proprietary information, such as full reasoning that should stay hidden. The third is account and platform integrity, such as mass account creation or bypassing rate limits at large scale. The fourth is other new abuse paths that lead to real harm and have clear fixes.
Model content issues are out of scope. Rude answers or information that a search engine already shows do not qualify. The issue must be in an active OpenAI product, it must be reproducible, and the report must say how to fix it. Only the first report of an issue is rewarded.
You do not need an invitation. Sign in to Bugcrowd and submit through the program page. Use only your own test accounts and never attack agents run by real users. The program offers safe harbor for good faith research. Bugcrowd shows submissions on this program accepted as recently as 14 October 2026.
Quick answers
Are jailbreaks in scope?
No. Model content issues and general jailbreaks are out of scope. OpenAI runs separate private campaigns for some harm types, such as biorisk.
What about normal security bugs?
Those go to the separate OpenAI Security Bug Bounty. Reports can be moved between the two programs depending on scope.