Every layer is a real buyer, real function

Pre-deploy scanning of skills, MCP configs, npm packages

One malicious skill install = agent hijack. The postmark-mcp incident silently forwarded 15,000 emails/day for months before detection. Scan before you trust the code an agent is about to run.

768 rules · 97.2% Garak recall · 0.48% FP (enforce lane)

Runtime prompt / tool / output defense

Even audited skills can be prompt-injected at runtime via external content (web pages, emails, docs). And skills that pass static audit can still be compromised mid-session via tool-response poisoning. Static scan is necessary; runtime enforcement is sufficient.

50ms median · inline MCP proxy

Behavioral anomaly detection with a 3-layer AI funnel

Rules catch the 90% of attacks you have seen before. The other 10% need AI — but if you call a cloud LLM on every request, cost and latency explode. The funnel keeps P50 under 50ms, P99 under 5s, at 95% cheaper than naive "always call GPT" architectures.

90/7/3 funnel · 7-day baseline

Honeypot integrated in Guard daemon

Passive defense is half the story. Honeypots convert the attacker's actions into your intelligence: tactics, tools, timing, infrastructure. You learn without leaking real data. The detections crystallize into rules that protect everyone on the TC network.

Integrated in Guard · zero-config

Auto-block, alert, playbook execution

Detection without response is noise. Agent attacks move in seconds — by the time a SOC analyst reads the alert, credentials are already exfiltrated. The loop has to close without a human for the 90th-percentile case.

11 actions · threshold policy