NomosGuard: Fail-Closed Security Core Where the LLM Writes Evidence but a Rules Engine Decides
Emergency-Tutor-7 · reddit · 2026-11-23
Open-source NomosGuard (v0.7.11, Apache-2.0, pip installable) cleanly separates LLM from decision-making:
- The model only writes evidence: the LLM reads events, configs, CVE records and logs and outputs structured claims, each citing a reproducible fragment (log-line hash, config path, CVE id); claims without citations are dropped
- Hash-chained ledger: surviving claims go into an append-only SHA-256 hash-chained ledger
- Rules engine decides: plain logical inference (no randomness, no model) derives attack and access paths, and a policy gate returns ALLOW / ALERT / BLOCK with deny as the default (fail-closed)
- The LLM can explain decisions afterward and draft remediation, but if its explanation contradicts the chain, the chain wins
Ships with 1361 tests, a reproducible demo scenario, real log ingestion and a benchmark corpus. Author's honest limits: it's a reasoning core, not a scanner — no host discovery, it only reasons over events you feed it; Linux CI fails on any skipped test, Windows measures 99.64% coverage (fcntl locks and directory fsync can't execute there).
More from coding & agent
- Stanford's free RAG lecture goes beyond chunking: BM25, HNSW, and agents — stanfordnlp · 2026-11-23
- Don't trust your agent's 'verified' claim: 5 practical rules for AI self-verification — dotey · 2026-11-23
- 44-Minute 3D Scene: Blender MCP + ThreeJs + Self-Built Skill Codex Workflow — OpenAIDevs · 2026-11-23
- Where Do You Draw the Line on Agent Permissions? Devs Weigh In on MCP's All-or-Nothing Token Problem — anishtr4 · 2026-11-23
- Wrapflow Beta: A Remote MCP Server That Gives ChatGPT and Claude a Database, Backend and Hosting — SupPandaHugger · 2026-11-23
- Inkling-2-Mini hits 10.04x on KernelBench-Mega locally, beating Grok 4.7 and GPT-6.1 — ziqiao_ma · 2026-11-23