Skip to content

Hermes AI Agent Security Standard v1.0

Density:

As organizations transition from static Large Language Models (LLMs) to autonomous agentic systemsโ€”capable of browsing networks, executing code, mutating databases, and coordinating multi-agent swarmsโ€”traditional web application security frameworks (such as OWASP Top 10) prove insufficient.

The Hermes AI Agent Security Standard (HASS-1.0) establishes an empirical, verifiable taxonomy spanning the full agent execution lifecycle:

HASS = ฮฃ (w_i ร— S_i) [where ฮฃ w_i = 1.0]

The primary question answered by this standard is:

โ€œIf this agent is compromised via indirect prompt injection or context poisoning, what can the adversary actually reach, modify, and exfiltrate?โ€œ


The standard partitions the agent attack surface into 7 discrete, auditable layers:

Layer IDNameWeight (w_i)Scope & Defense Focus
L1Model Security10%Foundation weights, adversarial prompt robustness, RLHF alignment, tokenization evasion, model integrity.
L2Agent Runtime Security15%ReAct loops, planning horizons, execution sandboxing, memory isolation, context injection filtering.
L3Tool Security15%Parameter schema validation, code interpreter constraints, network egress filters, destructive command gates.
L4MCP & Integration Security20%Model Context Protocol transport auth, server identity attestation, scope least-privilege, rogue tool injection.
L5Data & Memory Security15%Vector store poisoning (RAG), persistent episodic memory tampering, credential leakage in context windows.
L6Identity & Permission Security15%Non-human identities (NHI), ephemeral capability tokens, privilege escalation, machine credential rotation.
L7Infrastructure & Host Security10%Container/VM isolation, seccomp/eBPF confinement, kernel attack surface, Kubernetes network policies.

The Hermes Agent Security Score (HASS) produces a normalized value between 0 and 100, categorized into 4 operational risk tiers:

[0 โ”€โ”€โ”€โ”€โ”€โ”€ 39] CRITICAL RISK โ†’ Immediate production intervention required (T0)
[40 โ”€โ”€โ”€โ”€โ”€ 59] HIGH RISK โ†’ Exploitable lateral movement paths active (T1)
[60 โ”€โ”€โ”€โ”€โ”€ 79] MODERATE RISK โ†’ Standard enterprise deployment with residual risk
[80 โ”€โ”€โ”€โ”€ 100] HARDENED โ†’ Defense-in-depth, strict sandboxing, verified containment

The global score is subjected to non-linear penalties when critical architectural anti-patterns are observed:

  • Unrestricted Code Execution Tool: -25 points penalty if an agent can execute arbitrary shell scripts without air-gapped sandboxing.
  • Unauthenticated MCP Transport: -20 points penalty if local IPC or SSE/HTTP endpoints operate without mutual TLS or HMAC tokens.
  • Unbounded Memory Persistence: -15 points penalty if untrusted third-party inputs are indexed directly into shared RAG databases.
  • Absence of Kill-Switch (Out-of-Band): Caps maximum achievable score at 49/100 regardless of other controls.

HASS-1.0 mandates a deterministic evaluation of Blast Radiusโ€”the maximum perimeter reachable by an attacker assuming complete control of the primary agent loop:

[Indirect Prompt Injection]
โ†“
[Agent Reasoning Loop]
โ†“
[MCP Tool Gateway]
โ†“
โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ–ผ โ–ผ โ–ผ
[Internal Filesystem] [Production Database] [Cloud APIs / Secrets]
  1. CONTAINED (Level 1): Agent is restricted to ephemeral, read-only scratch space; tool calls have zero persistent side effects.
  2. LOCAL (Level 2): Agent can modify local application state but lacks network egress or privileged database access.
  3. LATERAL (Level 3): Agent holds credentials allowing inter-service communication, secondary agent invocation, or corporate email dispatch.
  4. CATASTROPHIC (Level 4): Agent possesses administrative credentials, write access to production databases, or host root capability.

Every compliance assessment against HASS-1.0 evaluates 35 verification controls divided across 5 control categories:

  1. Control Gate A โ€” Autonomous Action Limits: Verify that every state-altering action (financial transactions, database drops, email broadcasts) requires an authenticated Human-In-The-Loop (HITL) cryptographic approval gate.
  2. Control Gate B โ€” Ephemeral Token Exchange: Verify that long-lived API keys are never stored in prompt context; tools must receive just-in-time (JIT) capability tokens scoped to a single invocation with lifetime โ‰ค 60 seconds.
  3. Control Gate C โ€” MCP Strict Schema Validation: Enforce JSON schema validation and semantic sanitization on all MCP tool parameters before passing arguments to host runtimes.
  4. Control Gate D โ€” RAG Partitioning & Origin Tagging: Ensure ingested RAG chunks carry immutable provenance tags and cannot overwrite core system prompt directives.
  5. Control Gate E โ€” Out-of-Band Hardware Kill-Switch: Maintain an external watchdog capable of terminating agent worker processes and revoking cloud credentials within 500 ms.

The Hermes Codex platform provides turnkey implementations to test and audit environments against HASS-1.0:

Interactive Assessment Engine

Evaluate your agent stack client-side across all 7 layers in less than 3 minutes. Run Free Assessment โ†’

Agent Blast Radius Simulator

Simulate worst-case propagation paths when an agent tool or memory store is poisoned. Explore Blast Radius โ†’

Agent Threat Studio

Inspect weaponized exploit chains and defense profiles for autonomous architectures. Open Threat Studio โ†’

Commercial Audit & Certification

Commission a formal HASS-1.0 assessment report with executive roadmap and attestation. View Audit Services โ†’