If you find a security issue in OpenWar, please do not open a public GitHub issue.

Instead, report it privately:

• Open a private security advisory on this repo: https://github.com/PythonLuvr/openwar/security/advisories/new
• Or DM the maintainer on Discord: https://discord.gg/ku6GJS92V2

You can expect an initial response within 7 days. Confirmed issues will be patched in the next minor release, or sooner if the severity warrants it.

OpenWar is a runtime that drives third-party LLMs and tools. The security boundaries we maintain:

• Sandbox escape: a brief's tool calls should never reach paths or hosts outside the authorized scope.
• Authorization bypass: a brief's authorized_costs list should fully gate destructive operations. Bypassing it via tool name spoofing, category aliasing, or coordinator state manipulation is in scope.
• State corruption: the session state files under ~/.openwar/ should not be readable or writable by other users on the system.
• MCP transport: the JSON-RPC transport for MCP servers should validate handshake responses and reject malformed messages.

Out of scope:

• Vulnerabilities in third-party LLM providers (Anthropic, OpenAI, Gemini, Grok, etc).
• Vulnerabilities in MCP servers written by other authors.
• Prompt-injection attacks against the LLM itself (these are an active research area, not a runtime security boundary).

Pin to a known-good version in your package.json if you need long-term stability.