IRAM OS ships an MCP server that lets AI agents act on a desktop, so security reports are taken seriously. The policy boundary lives in bin/iram_policy.py and the permission tiers are described in the README.

Only the latest release and the main branch receive fixes.

Please do not open a public issue for a vulnerability. Use GitHub's private reporting instead:

1. Go to the Security tab of this repository.
2. Choose Report a vulnerability and describe the problem, affected component, and steps to reproduce.

Useful details: the output of iram doctor, the MCP permission tier you were running (read, desktop, privileged), and whether the issue needs local access or a malicious web page or model output.

This is a volunteer-run project, so there is no fixed response SLA. Reports are acknowledged on a best-effort basis, and credit is given in the fix unless you prefer otherwise.

• Bypasses of the MCP tier policy or argument validation (for example reaching exec, shell injection, path traversal)
• Secrets leaking into logs, ps output, or screenshots
• Unsafe archive handling in iram snapshot restore
• The Hub or agent daemon accepting browser-forged requests

• Anything that requires you to run the unrestricted command shown by iram-ai-cmd
• The optional FastMCP SSE service, which is documented as unauthenticated and off by default
• Vulnerabilities in upstream projects (Hyprland, Waybar, Ollama, Ubuntu); please report those upstream