If you discover a security vulnerability in d3vsh0p, please open an issue on GitHub.

Please include:

• A description of the vulnerability
• Steps to reproduce
• Potential impact
• Suggested fix (if any)

This policy covers the d3vsh0p orchestrator platform (platform/), agent templates (templates/), and CLI tooling. It does not cover projects built by d3vsh0p agents, which live in separate repositories.

d3vsh0p orchestrates AI agents (via the Claude Agent SDK) that have full tool access within project directories. The orchestrator runs agents with bypassPermissions enabled — agents can read, write, and execute commands without interactive approval.

What this means:

• Implementation agents have unrestricted access within the project directory
• The principal engineer (Morgan) reviews all changes before merge
• The PM agent (Riley) writes are sandboxed to the project directory via SDK PreToolUse hooks
• The orchestrator itself runs as the local user — it has whatever filesystem and network access the user has

Mitigations:

• Projects live in isolated repositories outside the d3vsh0p directory
• Agent prompts instruct agents to operate only within their assigned project
• Git worktrees provide filesystem isolation between parallel agents
• Health checks gate the session — broken projects are caught before agents start looping
• All agent work goes through code review (Morgan) before merge

This is a local development tool, not a multi-tenant service. It trusts the local user and the Claude API.