OpenAI Introduces Safety Bug Bounty Program

OpenAI has unveiled the Safety Bug Bounty program to prevent misuse and safety risks arising from the rapid advancement of AI technology. This program complements the existing Security Bug Bounty, addressing issues that may pose real safety threats even if they fall short of traditional security vulnerability criteria.

The main focus areas are as follows:

• Agentic Risks (including MCP): Agent hijacking and data exfiltration via prompt injection, large-scale unauthorized agent actions, etc.
• OpenAI Proprietary Information: Vulnerabilities exposing information related to model reasoning or other proprietary data.
• Account and Platform Integrity: Bypassing anti-automation controls, manipulating account trust signals, evading blocks, etc.

Simple Jailbreaks, the model's use of impolite language, exposure of information easily found through search, and other content policy violations without demonstrated real harm are excluded from rewards. However, for certain risks such as Biorisk, researchers will be recruited separately through private campaigns.