SPILLWAY defends tool-using agents from prompt injection
SPILLWAY separates target admission from sanitization using tool output fields and fixed rules, achieving zero successful injections across backbones and task suites.
Stories per day · 30 days
2 stories, newest first · one card per story
Stories tagged #agent-defense
AgentShield fingerprints AI agents via behavioral telemetry
AgentShield detects autonomous agents through PRA-loop telemetry exposing differences from humans while mitigating Extinction-induced Policy Drift attacks.