KeepKeys is security-sensitive software. No version is described as unbreakable or as protection against a compromised user session, administrator, debugger, malicious approved executable, or physical compromise.

Use this repository’s Security → Report a vulnerability flow to open a private GitHub Security Advisory. Do not open a public issue for a suspected vulnerability.

Never include a real API key, password, token, private key, seed phrase, customer record, credential-bearing screenshot, credential-vault export, or transcript. Reproduce with a generated synthetic value.

Please include:

• affected version and commit;
• operating system, desktop/vault provider where applicable, and agent-client versions;
• preconditions and realistic impact;
• minimal reproduction using synthetic data;
• evidence of whether the value reached chat, tool input/output, logs, files, another process, the public internet, or an unintended tailnet device;
• a suggested fix if you have one.

Neorome aims to acknowledge a well-formed report within seven calendar days. Disclosure timing is coordinated after a fix and verification are available.

Good-faith, non-destructive research against your own local clone and synthetic credentials is welcome. Do not test Neorome infrastructure, other users, third-party accounts, production services, or real credentials. Do not publish an exploit before coordinated remediation.

Changes to native-vault queries, native dialogs, phone-intake routing, shared tool schemas, client adapters, command construction, environment handling, output capture/redaction, cache compilation, marketplace wiring, or permissions require:

1. a threat-model update;
2. deterministic regression coverage;
3. full ./scripts/check and ./scripts/test;
4. platform-native doctor round trips for macOS Keychain, Windows Credential Manager, and Linux Secret Service;
5. macOS, Windows, and Linux CI;
6. independent security review of the final diff.

Phone-intake changes also require deterministic identity, browser-session, origin, one-use, expiry, content-size, listener-cleanup, and no-Funnel tests. Use a generated value and a tailnet owned by the researcher.