Skip to content

Cross-Agent Semantic Intent Verification & Guardrail Arbitration

Density:

In naive multi-agent architectures, an agent receiving an adversarial payload in an ingested CVE advisory or ticket description can experience instruction drift and instruct downstream tool-calling agents to drop firewall tables, invalidate root certificates, or isolate legitimate domain controllers:

[Adversary CVE Advisory]
| (Indirect Prompt Injection)
v
[CTI Analyst Agent] ---> (Compromised Intent) ---> [Firewall Sentinel Agent]
|
(Drops critical production rules!)

Hermes prevents cascade propagation by embedding both the stated operational rationale vector u and the concrete proposed system action vector v into a dense semantic vector space:

Semantic Divergence ฮ” = 1 - cos(u, v) = 1 - (u ยท v) / (||u|| * ||v||)
  • ฮ” <= 0.18 : NORMAL. High semantic alignment. Rationale and concrete action match defensive intent. Proceed to quorum voting.
  • 0.18 < ฮ” <= 0.35 : SUSPICIOUS. Rationale and concrete action exhibit semantic divergence. Escalated to secondary independent Arbiter Agent (AGENT-ID-ARBITER-07).
  • ฮ” > 0.35 : CRITICAL ANOMALY. Action rejected immediately. Upstream agent quarantined and logged for prompt injection forensics.

+---------------------------------------------------------------+
| Proposed Mitigation Payload |
+---------------------------------------------------------------+
|
+---> Layer 1: Deterministic AST Policy Linter
| (Bar wildcard IP drops, root cert purges, DC isolation)
|
+---> Layer 2: Dense Semantic Vector Classifier
| (Cosine distance check: ฮ” <= 0.18)
|
+---> Layer 3: Secondary Independent LLM Arbiter (SEV-SNP)
| (Runs isolated prompt injection detection prompt)
v
[ALL 3 CHECKS PASS] ===> Forward to Quorum Consensus (Section 71)
[ANY CHECK FAILS] ===> ABORT TRANSACTION & ALERT SOC

The entire 3-layer semantic intent arbitration pipeline executes within a strict 35-millisecond budget, preserving sub-second defense responsiveness.