AgentPolicyForge builds benchmarks and trains agent policy guards
Original title: AgentPolicyForge: From Executable Environments to Policy Benchmarks and Adaptive Guards
On PolicyGym, a trained guard cut attack success 40-44% relative to prompting and rejected fewer benign requests.
Researchers introduce AgentPolicyForge, which uses executable environments to build policy benchmarks and train guards for customized agent deployments. PolicyGym covers ten policy families, with 698 scenarios for training data and 254 held-out, human-reviewed test scenarios. Violations use permitted tools, so guards must check arguments, resources, and context. The authors report a 40-44% relative drop in attack success versus policy prompting alone.
Unconfirmed: The attack-success reduction is the authors' reported result on PolicyGym, not an independent evaluation.
- Training data come from 698 scenarios; tests use 254 held-out scenarios.
- Every violating case calls a permitted tool, so tool identity is not enough.
- Updating the strategy archive further reduced in-domain attack success.
- Guards are fine-tuned on balanced, environment-verified examples.
From the original article
1 min read at arxiv.orgCustomized agent deployments define policies over tools, resources, and execution conditions. Providing these policies in the task agent's prompt does not reliably prevent violations. Training a guard requires examples of tool calls that comply with or violate those policies.
By Chao Shang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.