Bug Bounty Researcher Details Multi-Turn AI Guardrail Bypasses

Summary

A bug bounty researcher describes three multi-turn methods for bypassing frontier-model safeguards, including study-guide framing, guessing games and string manipulation. The interview also covers Parseltongue and an in-progress practice playground.

Key points

• Dustin describes testing multi-turn guardrail bypasses in frontier models, including models designed to restrict CBRNE-related outputs.
• The examples layer study-guide framing, guessing games and persona or string-manipulation approaches to get models to produce restricted content.
• Dustin says a roughly $51,000 bounty involved his work, Edward's contributions and work from Toxik.
• Parseltongue is presented as an online, open-source tool for splitting prompts into variables and transforming them with different encodings for AI red teaming.
• Dustin says contributions to the Parseltongue repository carry supply-chain risks.
• An LLM practice playground with challenges was in progress; Dustin said it was intended to offer a bug-bounty- or appsec-like setting for trying multi-turn prompts.