If you discover a security vulnerability, please report it responsibly:

1. Do not open a public issue.
2. Use GitHub Private Vulnerability Reporting: Security -> Advisories -> New draft security advisory
3. Include a description, reproduction steps, and potential impact.

If private vulnerability reporting is not enabled in this repository, contact the maintainer through GitHub directly and do not publish details in a public issue.

open-compute drives a computer through an LLM. Treat it as you would any tool that can move the mouse, type, and launch applications:

• Run real backends in isolation. Use a dedicated VM or container, not your primary desktop. The library cannot sandbox the host for you.
• Keep a human in the loop. The default SafetyPolicy mode is confirm, which blocks risky actions (clicks, typing, key presses, app launches) unless a confirmation callback approves them. Do not switch to allow_all outside a disposable, isolated environment.
• On-screen content is untrusted input. A page or app the agent views can contain prompt-injection text. The model may be steered by it; the safety gate and human confirmation are your defense.
• Secrets. Never hard-code API keys. The Claude/OpenAI backends read keys from the environment via their SDKs. Do not log screenshots or transcripts that may contain credentials.

For smaller solo projects, response times may vary. Critical issues will be prioritised. Please allow reasonable time before public disclosure.