Resources
Practical guides on keeping AI coding agents inside their boundaries: what they read, what they run, and which credentials they can reach. Checked against each vendor's own documentation.
Field guides for the teams putting Claude Code, Codex, Cursor and MCP servers in front of real code.
Start here
Permissions, managed settings, sandboxing, hooks and MCP approval for Claude Code, plus the gaps a team has to close itself.
26 October 202611 min readGuide
26 October 20269 min readGuide
26 October 20267 min readGuide
26 October 20267 min readGuide
26 October 20267 min readGuide
26 October 20267 min readGuide
26 October 20265 min readGuide
26 October 20266 min readGuide
Coverage
A coding agent's request crosses four boundaries before it takes effect. Read across a row to see what each guide covers.
| Guide | WorkspaceWhat can enter the agent's view? | ContextWhat can steer its decisions? | AuthorityWhat can actually take effect? | EvidenceWhat can the team prove later? |
|---|---|---|---|---|
| Claude Code securityWorkspaceAuthorityEvidence | Covered | Not covered | Covered | Covered |
| MCP securityContextAuthorityEvidence | Not covered | Covered | Covered | Covered |
| AI coding agent securityWorkspaceContextAuthorityEvidence | Covered | Covered | Covered | Covered |
| Prompt injection protection for coding agentsContextAuthority | Not covered | Covered | Covered | Not covered |
| AI agent runtime securityAuthorityEvidence | Not covered | Not covered | Covered | Covered |
| Codex securityWorkspaceAuthority | Covered | Not covered | Covered | Not covered |
| Cursor securityWorkspaceContextAuthority | Covered | Covered | Covered | Not covered |
| How to keep API keys and .env files away from AI coding agentsWorkspaceContext | Covered | Covered | Not covered | Not covered |