PandaGuard: LLM Safety Evaluation Framework

PandaGuard models LLM safety assessment as interactions between attacker, defender and judge agents, enabling reproducible evaluation of a wide range of jailbreak attack methods and defense mechanisms across different LLM back-ends. The associated PandaBench benchmark aggregates large-scale evaluation data on 40+ models, 19 types of attacks and 12 defense strategies, providing a standardized way to compare vulnerabilities, study trade-offs between safety and utility, and inform Beijing-AISI’s research on robust and governable language models.

Endorsements support Beijing Institute of AI Safety and Governance (Beijing-AISI).

No endorsements yet.

Discussion

Start Date

-

End Date

-

Expected Duration

-

Funding Raised to Date

-