The authors introduce AgentPolicyForge, a framework that binds deployment policies to concrete tool calls and executes them in resettable environments.
Executable environments enable policy benchmarks and iterative guard training
AgentPolicyForge constructs the PolicyGym dataset and co-evolves attack strategies with guard fine-tuning, reducing violations by 40–44% over policy prompting alone.
Big Tech
Chao Shang · Wanyu Du · Yi Zhang · Sandesh Swamy
AWS AI Labs
Research Digest··3 min read
The authors present AgentPolicyForge, a framework that uses executable environments to create policy-compliance benchmarks and iteratively train policy guards through coevolution of attacks and defenses.
§