Report AI agent safety flaws for OpenAI's Safety Bug Bounty
Sign up on Bugcrowd and read OpenAI's Safety Bug Bounty brief. It pays for AI-specific abuse and safety issues that don't count as normal security bugs: for example a third-party prompt injection that hijacks ChatGPT Agent, Atlas or Codex into a harmful action or a data leak, agentic products doing disallowed actions at scale, leaks of hidden reasoning, or ways to get around account restrictions and anti-automation controls. Plain jailbreaks and rude-content bypasses are out of scope. Test only with accounts you own, never against real users' agents, and follow third parties' terms when testing MCP and connectors. A prompt-injection report must reproduce at least half the time, and every report needs clear steps and a suggested fix. Only the first report of an issue is paid.
Details
- How it works
- Sign up on Bugcrowd and read OpenAI's Safety Bug Bounty brief. It pays for AI-specific abuse and safety issues that don't count as normal security bugs: for example a third-party prompt injection that hijacks ChatGPT Agent, Atlas or Codex into a harmful action or a data leak, agentic products doing disallowed actions at scale, leaks of hidden reasoning, or ways to get around account restrictions and anti-automation controls. Plain jailbreaks and rude-content bypasses are out of scope. Test only with accounts you own, never against real users' agents, and follow third parties' terms when testing MCP and connectors. A prompt-injection report must reproduce at least half the time, and every report needs clear steps and a suggested fix. Only the first report of an issue is paid.
- Tools
- Bugcrowd (OpenAI Safety engagement), ChatGPT, Codex or Atlas on your own test accounts, A screen recorder and notes for reproduction steps. Tool links
- Startup costs
- $0 to join (testing some agentic products may need a paid ChatGPT plan, which isn't reimbursed)
- Reported earnings
- Per OpenAI's announcement and SecurityWeek, accepted reports earn up to $7,500. Third-party trackers of the Bugcrowd table list P1 $5,500 to $7,500, P2 $2,500 to $3,500, P3 $750 to $1,500 and P4 $250 to $500, and one tracker counted only 11 rewarded reports since July 2025 (average $812.50). OpenAI sets the final amount.The $7,500 maximum is from OpenAI's own announcement; the tier ranges, report count and average award come from a third-party tracker of the Bugcrowd page.Entry last updated · What the labels mean
Sources
A full research note (hours, startup costs and a month-by-month earnings range) for this entry is in progress.
Straight answers
Related entries
AI data annotation / model evaluation
Rate and correct AI model answers on annotation platforms, paid hourly
| Startup cost | $0 |
|---|---|
| First dollar | 1–6 weeks after applying |
| Month 12 | $300–$2,000/mo |
Record chore videos to train household robots
Film first-person chores for robot-training firms; reported $3–$25/hr
| Startup cost | $0–$30 |
|---|---|
| First dollar | 1–3 weeks |
| Month 12 | $50–$300/mo |
User testing of AI products
Test sites, apps and AI features for pay; moderated interviews pay $60–$100/hr
| Startup cost | $0 |
|---|---|
| First dollar | 1–3 weeks after joining |
| Month 12 | $100–$500/mo |