Security and access control for AI coding agents
Agent permissions
Permissions and least privilege for agents
- A .agentignore for Your Agent: What to Hide and Leave Visible
- A File-Access Policy That Survives a Repo Move
- Auditing an Allowlist: Finding Rules No Agent Fires
- Guardrails for Legacy Interfaces: What an Agent May Not Touch
- One Allowlist for Tool Calls and File Reads
- Read-Only by Default: Agent Tasks Without Writes
- What an Agent May Do to a Database: Grants, Roles, Read-Only
- A Deny List That Outranks the Allowlist: Precedence and Tests
- All of Agent permissions
Agent misbehavior
Detecting and blocking agent misbehavior
- Blocking an Agent from Editing Its Own Checks
- Detecting an Agent That Retries Its Way Around a Limit
- Detecting an Agent That Rewrites the Prompt to Make Its Test Pass
- Instructions Hidden in a Pull Request an Agent Reads
- Prompt Injection Through Rows an Agent Reads
- What a Pre-Commit Hook May Block and What It Must Only Warn
Audit logging
Audit logging for agent actions
- A Task Log for Agent Work: What Was Asked and What Shipped
- Auditing Human Decisions in the Task Log
- A Fallback Evidence Bundle When the Link Dies
- A Tamper Log Schema: Fields, Joins and Alerts
- A Worktree Naming Scheme That Keeps Traces Aligned
- Agent Commit Messages as a Search Key for Later Audits
- Approval Logs That Join Proposals to Applied Rows
- Audit Logs for Agent Actions: What to Record and How to Query
- All of Audit logging
Agent sandboxing
Sandboxing and isolating agent runs
- Running Agents on Ephemeral CI Runners Instead of Your Laptop
- Using Git Worktrees to Isolate Agent Changes
- A Pre-Edit Gate That Cannot Write: Scratch Clones and Deny Paths
- A Sandbox That Refuses Irreversible Calls
- Blocking Agent Writes That Escape the Worktree
- Blocking Install Scripts in Agent Environments
- Cleanup Rules for Worktrees: When to Prune, Lock or Archive
- Denying Package Install Scripts in Agent Sandboxes
- All of Agent sandboxing
Agent secrets
Secure secrets and signing for agents
- Signed Agent Commits: Key Scoping and Verification
- A Job-Level Secret Broker for Agents: Scoping Credentials
- Auditing a Context File for Injection and Secrets
- Auditing Who Can Mint an Agent's Signing Key: A Roles Table
- Designing an Incident Runbook for a Leaked Agent Credential
- Redacting Sensitive Fields Before Output Reaches a Model
- Revoking a Key an Agent Could Read: A Runbook and Its Timers
- Revoking a Package Registry Token an Agent Can Read
- All of Agent secrets