OpenAI launches Safety Bug Bounty for AI risks | Keryc
OpenAI announces a public program called Safety Bug Bounty to identify abuses and security risks in its AI products. Why does that matter to you? Because AI moves fast, and with it come new, concrete ways to cause harm. This program aims to close that gap by collaborating with researchers and ethical hackers.
What the program looks for
The focus is specific: the program accepts reports that involve real risks of abuse and security, even when they aren’t traditional software vulnerabilities. OpenAI complements its Security Bug Bounty program and routes investigations between the two teams as appropriate.
Scenarios in scope include:
Agentive risks including MCP: when malicious text manages to take control of an agentic agent (for example Browser or ChatGPT Agent) to perform harmful actions or leak sensitive data. The behavior must be reproducible in at least 50% of attempts.
Unauthorized actions at scale carried out by an agentive product on OpenAI’s site.
Potentially harmful actions by agentive products not explicitly listed, as long as there is evidence of plausible, material harm.
Model generations that reveal proprietary information related to reasoning or vulnerabilities that expose other proprietary OpenAI data.
Account integrity and platform signal issues: for example, bypassing anti-automation controls, manipulating account trust signals, or evading suspensions and restrictions.
What's out and warnings
General jailbreaks that only produce rude language or responses easily found on search engines are out of scope.
Tests for some risks (like MCP) must respect third-party terms of service.
For certain harmful categories, OpenAI runs private campaigns (for example, bio-risk issues in ChatGPT Agent and GPT-5) and interested researchers can apply when those calls are announced.
Not every policy bypass is eligible. What they’re looking for are paths that genuinely enable harm to users and that have clear mitigation steps.
How to participate if you're a researcher
Apply through OpenAI’s Safety Bug Bounty form.
Provide reproducible evidence and, when applicable, plausible impact and concrete mitigation steps.
Keep in mind reports will be triaged by Safety and Security teams and may be rerouted depending on scope and responsibility.
Practical impact and why it should matter to you
If you work with AI products (as a developer, integrator, or security lead), this means there’s now a formal channel to report specific agentive behavior and abuse risks that could previously fall into a gray area. For users and businesses, it signals that OpenAI is strengthening defenses beyond classic security patches: they’re also working on the safety of model behavior.
Curious but not a practitioner? This shows something important: AI safety isn’t just code — it’s design, controls, and community. Incentives to report failures help make tools safer for everyone.
Final reflection
The Safety Bug Bounty is a practical step: it acknowledges that AI introduces new vectors of risk and that the research community is key to spotting them. It’s not a silver bullet, but it’s an invitation to collaborate responsibly to reduce real harms.
