article methodology

Prime Agent security: trust boundaries and the non-sandbox warning

Prime Agent executes model-generated Python, shell commands, skills, and extensions with your OS permissions. Worker and kernel process separation is for lifecycle isolation, not security isolation. Use disposable clones, clean worktrees, and external restricted environments for untrusted code or instructions.

v v0.7.1reviewed 2026-08-09evidence officialsources S003, S008, S011, S015, S021cutoff 2026-08-09
Home

First-party announcement, documentation, release, or installer.

Trust model

Prime Agent is best understood as a local trust boundary expansion, not a security sandbox. The model can steer code, but it does not gain magical confinement around that code.

Kernel and workers

Kernel and worker separation keeps failure domains cleaner, but it does not stop file access, shell access, or credential use within your account.

Installer review

Review the installer if you care about supply-chain risk. The shell-pipe install is convenient, but convenience and strict review are in tension.

Skill and extension trust

Skills, extensions, and packages are trusted code. If you cannot audit them, do not grant them sensitive workspaces or credentials.

Transcripts and secrets

Session exports and transcripts can leak secrets. Redact before sharing, and prefer disposable environments for sensitive projects.

Unresolved questions

  • prompt injection persistence,
  • trustworthy auto-install boundaries,
  • and safe default policies for untrusted repos.