Best Way to Audit What Permissions Your AI Tools Actually Have

[DarkLantern]

Following the Miasma supply chain attack that targeted AI coding tools through their configuration systems, a lot of people are realising they are not certain what their AI assistants can actually access.

How do you check what permissions Claude Code, Copilot, Cursor, Gemini CLI or any other AI coding tool has been granted? Where do these tools store their config files? What should you be looking for? Specific, actionable responses preferred.

[PlanetOftheApes]

On Mac, Claude Code stores its config in ~/.claude/. The settings.json file in that directory is what the Miasma attack targeted. Check that file for unexpected hooks in the SessionStart or other event fields

[EntangledOne]

For VS Code extensions generally, check the extension's requested permissions in the marketplace listing. Most AI coding extensions request far broader access than they need to justify

[NatureBoyRyan65]

The key thing to audit is whether your coding assistant has access to your credential stores, SSH keys, or environment variables containing API keys. That is what Miasma was harvesting

[Debbie96]

Run your AI coding tool in a dedicated user account with limited permissions if your threat model is high. The capability loss is real but the isolation is genuine

[WarningPoint49]

The simplest practical step: never open a repository in an AI coding tool without checking it for unexpected config files first. A quick check on the root takes thirty seconds

[Robin13]

Set your coding tool's working directory explicitly to the project folder rather than your home directory. This limits what it can traverse

[Ryan65]

Check your CI/CD pipeline configs for any references to AI tool config files that might pull from external sources