OpenHand executes AI-generated actions against real systems (the filesystem, the shell, the network, email). We take security reports seriously.

We follow a "current minor + previous minor" support window during the 0.x series. Once a 1.x release ships, the policy formalises into "current minor plus the previous minor for six months."

Pre-1.0 minors do not guarantee API compatibility; we will document any breaking change in CHANGELOG.md. After 1.0, the previous minor line will receive security fixes for six months from the date the next minor ships.

Please do not open a public GitHub issue for security problems.

Use one of the following private channels:

1. GitHub Security Advisories (preferred): https://github.com/ricardo-foundry/openhand/security/advisories/new
2. Email: security@openhand.dev (PGP key available on request).

Please include:

• A description of the issue and its impact.
• A minimal reproduction (commit SHA, config, repro steps).
• Your assessment of severity (sandbox escape, RCE, data exfiltration, etc.).
• Whether you would like public credit once the fix ships.

In scope:

• Sandbox escape from packages/sandbox (shell, filesystem, network).
• Prompt-injection that causes OpenHand to execute unapproved tools.
• Authentication / authorization flaws in apps/server.
• Secret leakage through logs, errors, or the web UI.
• Supply-chain issues in published packages.

Out of scope:

• Vulnerabilities in third-party LLM providers themselves.
• Attacks requiring a compromised machine or a malicious user with full shell access already (we are not an anti-malware product).
• Denial of service via deliberately expensive LLM prompts.

We will not pursue legal action against researchers who:

• Make a good-faith effort to avoid privacy violations and service disruption.
• Give us a reasonable window to remediate before disclosure.
• Do not exfiltrate more data than necessary to demonstrate the issue.

Thank you for helping keep OpenHand and its users safe.