Case Studies◆ AI-generated · Sourced
Introducing the OpenAI Safety Bug Bounty Program
TL;DR
OpenAI has launched the Safety Bug Bounty Program to solicit reproducible vulnerability reports against GPT-4, GPT-5.5, and related models—specifically targeting agentic vulnerabilities, prompt injection, and data exfiltration.
Background and Scope
The OpenAI Safety Bug Bounty Program is a core component of OpenAI’s operational safety framework, covering production models (e.g., GPT-4) and experimental releases (e.g., GPT-5.5 Bio), with explicit emphasis on real-world abuse vectors and deployment-level failure modes.
Target Vulnerability Classes
- Agentic vulnerabilities: Unauthorized tool use, goal misgeneralization, or unbounded planning behavior in autonomous agent workflows;
- Prompt injection: Direct or indirect injection leading to system prompt override, role hijacking, or context poisoning;
- Data exfiltration: Leakage of training data fragments, user session artifacts, or sensitive metadata via model outputs.
Operational Principles
- Reports must include full reproduction steps, impact chain analysis, and a minimal PoC;
- Theoretical risks or non-triggering configuration issues are out of scope;
- Bounties are tiered by severity (Critical/High/Medium), with maximum rewards up to $20,000 USD;
- The GPT-5.5 Bio Bounty is a specialized track under this program, focused on domain-specific risks in biomedical AI—e.g., hallucinated clinical recommendations or misinterpreted genomic sequence semantics.