PragLocker: Protecting Agent Prompts from Unauthorized Reuse
A team of researchers has introduced PragLocker, a strategy designed to safeguard LLM agent prompts as intellectual property in environments lacking trust. This innovative system tackles four primary issues: proactivity, runtime security, usability, and non-portability. By embedding semantics within code symbols and adding noise informed by target-model feedback, PragLocker generates function-preserving obfuscated prompts, ensuring they operate exclusively with the designated LLM. Tests conducted across various agent systems, datasets, and foundational LLMs demonstrate that PragLocker significantly limits cross-LLM portability while preserving target performance and resilience against adaptive threats. This research was published on arXiv with the identifier 2605.05974.
Key facts
- PragLocker is a prompt protection scheme for LLM agents.
- It addresses proactivity, runtime protection, usability, and non-portability.
- It constructs obfuscated prompts using code symbols and target-model feedback.
- Experiments show reduced cross-LLM portability and maintained target performance.
- Published on arXiv with ID 2605.05974.
- The research targets untrusted deployments of LLM agents.
- Adversaries can copy prompts for use with other proprietary LLMs.
- The scheme is robust against adaptive attackers.
Entities
Institutions
- arXiv