OpenAI Launches AI Safety Bug Bounty to Combat Prompt Injection and Agent Abuse

OpenAI is expanding its efforts to secure AI systems with a new Safety Bug Bounty program, targeting emerging risks tied to agent behavior and misuse. The move comes shortly after the company introduced features like the ChatGPT Library and in-app shopping, signaling a broader push toward real-world AI integration.

Focus shifts from traditional security to AI behavior risks

Unlike traditional security programs, this new initiative focuses on non-conventional threats such as prompt injection, data exfiltration, and harmful actions carried out by autonomous AI agents. As AI systems gain more capabilities, these risks have become harder to detect using standard vulnerability testing.

What the safety bug bounty program covers

The program complements OpenAI’s existing Security Bug Bounty but shifts attention toward how AI behaves in real-world scenarios. Researchers must demonstrate reproducible issues, with bugs expected to occur roughly 50% of the time to qualify.

OpenAI is also opening the door to reports involving exposure of sensitive internal information, including proprietary model behavior or reasoning patterns. Vulnerabilities that affect platform integrity, such as bypassing safeguards or restrictions, also fall within scope.

What is not eligible for rewards

However, not everything qualifies. Simple jailbreaks or low-impact content policy bypasses are excluded, along with reports that lack clear real-world harm or actionable fixes. The company aims to prioritize meaningful, high-impact findings over theoretical or minor issues.

Review process and participation requirements

Submissions will undergo review by both safety and security teams, with some reports potentially redirected depending on their nature. OpenAI also noted that certain edge cases may still qualify if they demonstrate credible harm.

The program is open to ethical hackers, researchers, and security professionals, though participants must apply through official channels. OpenAI may also introduce private bounty initiatives focused on specific high-risk areas in the future.

AI safety becomes a new cybersecurity frontier

This launch highlights a growing shift in the AI industry, where traditional cybersecurity models no longer fully address the risks posed by intelligent systems. Instead, companies are beginning to treat AI behavior itself as a new attack surface.

OpenAI expands ecosystem while cutting underperforming tools

Meanwhile, OpenAI continues to evolve its broader ecosystem. The company is reportedly working on a “super app” that would unify ChatGPT with other OpenAI services. At the same time, it has started trimming less successful offerings, including the shutdown of its Sora AI video tool, as it refocuses on scalable and commercially viable products.