AgentPolicyForge builds benchmarks and trains agent policy guards
On PolicyGym, a trained guard cut attack success 40-44% relative to prompting and rejected fewer benign requests.
Stories per day · 30 days
1 stories, newest first · one card per story