Security and Permissions
This page is the baseline policy starter for AI coding tool access.
Policy Objective
Reduce the risk of source-code leakage, destructive automation, and unsafe privilege sprawl when using coding agents or MCP integrations.
Scope
- local repository access
- shell execution
- network access
- MCP integrations
- prompts that can modify systems
Audience
- all contributors
- reviewers
- security approvers
Definitions
Least privilege: minimum access required for the taskSensitive repo: repo containing private code, secrets, regulated data, or production infrastructureDestructive action: delete, deploy, revoke, merge, publish, or data-changing operation
Mandatory Requirements
- tools must start with the least access needed
- destructive actions require human review
- production credentials must not be pasted into prompts
- write-enabled MCP servers require explicit approval
Recommended Controls
- use separate test accounts
- prefer read-only pilots
- keep repo instructions small and current
- log what validation commands were run
Roles and Ownership
- tool user: chooses the narrowest safe scope
- reviewer: confirms human-checkpoint requirements were followed
- security owner: approves higher-risk integrations
Approval Process
- read-only local repo access may be approved first
- shell, network, and write-enabled integrations require a higher review tier
- production-adjacent access requires a documented business reason
Exception Process
Every exception should record owner, duration, repository scope, credentials involved, and rollback plan.
Audit Evidence
- approved-tool matrix
- MCP evaluation record
- repo instruction files
- validation logs or PR evidence
Enforcement
Unsafe prompts, over-broad credentials, or unapproved integrations should be removed or disabled immediately.
Review Cadence
- review on every major tool or policy update
- at least quarterly for approved high-risk tools
Incident Response
- stop the tool or integration
- rotate exposed credentials
- inspect logs, prompts, and changed artifacts
- document the corrective action
Revocation Process
- remove access from the host
- revoke tokens or app grants
- remove unsafe shared prompts or config files
Copy-Ready Policy Template
All AI coding tools must use the least privilege required for the task.
Human review is mandatory before merge, deploy, destructive shell commands, permission changes, or write-enabled MCP actions.
Production credentials, customer data, and private secrets must not be supplied to prompts except through an approved secured workflow.
Adoption Checklist
- least-privilege default documented
- destructive actions gated
- credential rules published
- revocation path tested
Verification Note
Policy starter only. This page is not legal, compliance, or incident-response advice for every environment.