Skip to main content

Security Architecture & Threat Model

M31A is built on the principle that autonomous AI agents are untrusted actors operating inside your development environment.

Threat Assumptions​

  • Model proposals may contain hallucinations, invalid syntax, or unintended side effects.
  • Repository contents may contain prompt injection payloads designed to trick agents into exfiltrating secrets.
  • External dependencies or build scripts may contain malicious post-install hooks.

Defense Layers​

  1. Non-Bypassable Policy Gate: Evaluates every tool execution before invocation.
  2. Process Sandboxing: Landlock filesystem sandboxing and seccomp system call filtering.
  3. Sensitive Path Blocking: Absolute prohibition on accessing SSH keys, cloud tokens, or password stores.
  4. Secret Masking: Prevents credentials from ever appearing in logs, events, or UI projections.