NEOS Guard — OWASP LLM Top 10 Coverage

Mapped to the enterprise LLM-security evaluation checklist · honest status (2025)
3
Covered
3
Partial
4
Out of scope (honest)
Honesty principle: only what a runtime LLM firewall actually handles is marked 'covered'. Supply chain, data poisoning, embeddings and misinformation are separate product categories, so we honestly mark them 'out of scope' — that earns more trust than an inflated "covers everything" claim. Of the 6 runtime-defensible items: 3 fully covered + 3 partial.
IDThreatStatusNEOS Guard coverage
LLM01Prompt Injection
Novel 83% / known 99% detection · 0% FP · holdout-verified
✅ CoveredInput guard — rules + Meta Prompt Guard 2 + LLM ensemble, self-evolving
LLM02Sensitive Information Disclosure
Known 100% / novel 67% deterministic detection · 0% FP
✅ CoveredOutput guard — secrets engine (gitleaks-grade) + PII engine (Presidio-lite), self-evolving
LLM03Supply Chain
Honestly a different product category. NEOS Guard focuses on runtime defense
⚪ Out of scopeOut of scope — model/dependency supply-chain integrity is SCA/SBOM territory, not a runtime guard
LLM04Data & Model Poisoning
A runtime guard can't defend this. Marked honestly as not covered
⚪ Out of scopeOut of scope — training-data/model poisoning is training-pipeline governance
LLM05Improper Output Handling
Flags downstream-execution risk. Opt-in to avoid FPs for code assistants
🟡 PartialDangerous-output detection (opt-in) — SQL/XSS/shell/path-traversal/exfil links in output
LLM06Excessive Agency
Known 100% / novel 50% · blocks excessive actions via runtime session monitor
🟡 PartialBehavior guard — destructive actions, hijack execution, privilege escalation, session trajectory
LLM07System Prompt Leakage
Deterministic detection of system-prompt / internal-rule disclosure · 0% FP
✅ CoveredOutput guard prompt_leak + self-learned (internal-rule / hidden-rule disclosure)
LLM08Vector & Embedding Weaknesses
Honestly a different layer. Could extend to a RAG guard later
⚪ Out of scopeOut of scope — RAG embedding / vector-DB security is retrieval-infra territory
LLM09Misinformation
Not what a security guard does. Marked honestly as not covered
⚪ Out of scopeOut of scope — hallucination/factuality is fact-check/grounding product territory
LLM10Unbounded Consumption
Detects automated/DoS-style consumption patterns + per-tier rate limits
🟡 PartialBehavior guard (runaway loops · bot-speed) + API rate limiting
Numbers verified in the reproducible report · self-improvement in the evolution log