I built AgentInterdict to secure the boundary where an AI agent's context becomes action stopping prompt injection, credential leakage, memory poisoning, and authority spoofing before they can cause a breach. I benchmarked it against 200 real injection attempts and published the results honestly: 96.5% block rate, 7 documented misses (I publish the misses rather than claim 99.9% a security tool that hides its failures isn't trustworthy). The 7 misses are the gaps I care about most. They break down roughly as: - multi-turn / conversational injection that spans several exchanges - obfuscation techniques I haven't seen yet - edge cases in tool-call semantics So here's the question for this thread: what attack vectors or evasion techniques would you want an agent-security runtime to handle next? Whether you're building agents, running local LLMs, or thinking about this from a security perspective I want the hard cases. What am I missing? Open-source, local-first, free for personal/research use: github.com/BryanFiFife/AgentInterdict
AgentInterdict is runtime enforcement for autonomous AI agents — interdicting unsafe action by verifying authority, tracking provenance, and revalidating at action time.
