Safe REA Skill
A double-layer sandbox for open-source reverse engineering that contains REA and its receiving agent in separate restricted Docker containers.
0 projects
About Safe REA Skill
Safe REA Skill is a security-focused deployment framework for the REA (Reverse Engineer Anything) toolkit. It establishes a double-layer sandbox using Docker to isolate both the analysis tools and the AI agent interpreting the results. The environment is designed for static analysis of unfamiliar binaries, archives, and JavaScript bundles without host execution. ### Architecture and Interfaces The system utilizes a trusted Python controller on the host (Linux or macOS) to manage two restricted Docker containers. The first container runs REA with locked dependencies (including Ghidra and JADX), while the second container hosts the agent client (Claude Code or Codex). Communication is restricted to a bounded MCP channel and a fixed model provider endpoint, with the agent container having zero host mounts. ### Security Controls - **Network Isolation:** Both containers operate with `network=none` to prevent data exfiltration. - **Data Handling:** Only frozen, hash-approved copies of selected files are shared with the REA container via read-only mounts. - **Resource Limits:** Supports standard and large profiles, with the latter providing up to 5 GiB RAM and 4 CPUs for complex binaries. - **Verification:** Includes a suite of isolation checks and offline analysis tests to verify the boundary before live model interaction. ### Supported Environments - **Native Runtime:** Verified for POSIX-compliant Linux and macOS. - **Windows:** Support is currently experimental; while policy tests pass, full Docker Desktop integration is unverified. - **AI Clients:** Specifically guides setup for Claude Code and Codex using existing user subscriptions rather than API keys.
Summarised from the official page linked above.