Security is a primary concern for the AutoWeave library, as it serves as an execution and orchestration layer for agent-generated code and commands.

Currently, the main branch and the latest published version are supported for security updates.

• Sandbox Execution: AutoWeave does not provide a sandbox for unsafe code execution. It is purely the orchestration layer. Any remote workers (e.g. OpenHands) provisioned by tasks must be executed inside a hardened sandbox (e.g., Docker, gVisor) configured by the downstream product or user.
• Product Authentication: This library does not handle user authentication, OAuth, or RBAC. The downstream product shell is responsible for securing access to the orchestration engine.

• Trusted: The local filesystem (templates, bootstrap files), environment variables (.env.local), and configuration maps are considered trusted.
• Untrusted: Natural language requests passed to the compiler, and code/artifacts generated by LLM agents.

• Strict Validation: All state transitions and payload schemas are strictly validated using pydantic.
• No eval(): The library deliberately avoids unsafe dynamic code execution methods such as eval() or exec() internally.
• Secret Scanning: All commits to this repository are scanned for hardcoded secrets via gitleaks.
• Dependency Audits: Dependencies are continuously audited using pip-audit.
• Static Analysis: CodeQL runs on every pull request to analyze the Python source code for logic vulnerabilities.

If you discover a security vulnerability within AutoWeave Library, please DO NOT open a public issue.

Instead, please use the GitHub Security Advisory feature to privately report the issue:

1. Go to the Security Advisories tab for the repository.
2. Click "Report a vulnerability".
3. Provide full details of the vulnerability, including steps to reproduce.

• We will acknowledge receipt of your vulnerability report within 48 hours.
• We will provide a status update or resolution plan within 7 days.
• If a fix is required, we aim to release a patch within 14 days of the report.

All security vulnerabilities will be promptly addressed, and reporters will receive credit if they desire.