Skip to content

SECURITY.md

Latest commit

 

History

History
36 lines (26 loc) · 1.95 KB

SECURITY.md

File metadata and controls

36 lines (26 loc) · 1.95 KB
Raw

Security Policy

How MorningStar Works

MorningStar runs Claude Code CLI with --permission-mode dontAsk and full tool access (Read, Write, Edit, Bash, Glob, Grep) in your target repository. This means it can execute arbitrary shell commands and modify any file in the repo directory with the permissions of your OS user.

This is by design -- the agent needs to install dependencies, run tests, and write code. However, it means you should:

  • Only run MorningStar on repositories you trust
  • Only point it at Notion PRDs you trust (PRD content is passed to the LLM and could influence agent behavior)
  • Review generated commits before pushing to shared branches
  • Use the --dry-run flag to preview tasks before execution
  • Set conservative --budget limits
  • Run in an isolated environment (VM, container) for untrusted PRDs

Security Mitigations

  • Confirmation prompt: MorningStar asks for confirmation before executing tasks (bypass with --yes)
  • Budget limits: Per-task and total budget caps prevent runaway costs
  • Sensitive file exclusion: git add excludes .env, *.pem, *.key, credentials.json, and other secret patterns
  • Task ID sanitization: AI-generated task IDs are sanitized to prevent path traversal
  • Session ID validation: Claude session IDs are validated before reuse
  • Slack webhook validation: Only hooks.slack.com URLs are accepted
  • Model allowlist: Only known Claude model names are accepted
  • PRD fetch isolation: PRD fetching runs in a temp directory with read-only tools
  • No home directory access: Agent prompt explicitly prohibits reading ~/.ssh, ~/.aws, etc.

Reporting Vulnerabilities

If you discover a security vulnerability, please report it privately:

  1. Do not open a public GitHub issue
  2. Email: [security contact email]
  3. Include: description, reproduction steps, and potential impact

We will acknowledge within 48 hours and provide a fix timeline within 7 days.