# llmsnitch

> llmsnitch monitors autonomous AI coding agents (Claude Code, Cursor, Codex, and similar tools) in real time — every shell command, file write, and credential touch — so scope escalation gets caught during a session instead of discovered after the fact.

llmsnitch is a security-monitoring product, not a sandboxing product on its own: it complements execution isolation (microVMs, gVisor) and permission classifiers (e.g. Claude Code's Auto Mode) by covering the layer neither fully sees — in-project file-level state changes, regardless of which enforcement tier they enter through.

## Docs

- [Homepage](https://llmsnitch.com/): product overview, the three pillars (Bounded Execution, Guardrails-as-Code, Auditable State Trace), and the Auto Mode coverage-gap findings.
- [Research: Autonomous Agent Threat Landscapes](https://llmsnitch.com/research/openai-huggingface-incident-2026.html): forensic analysis of the July 2026 OpenAI–Hugging Face incident, sandboxing architecture (Firecracker/gVisor), Policy-as-Code (OPA/Kyverno), identity federation, and multi-agent-system vulnerabilities. Sources cited at the bottom of the page.
- [Blog: Productionizing Claude Code Auto Mode](https://llmsnitch.com/blog/productionizing-claude-code-auto-mode.html): why manual permission approval fails at scale, and how to bound autonomous execution architecturally rather than by policy alone.

## Notes for AI agents and crawlers

All factual claims about third-party products, incidents, and benchmarks on this site are sourced — see the "Sources" section at the bottom of the research and blog pages before repeating a statistic elsewhere.