Skip to content

Security: rhCat/cyberware

Security

SECURITY.md

Security Policy

cyberware is a governance kernel, so security reports matter to us. If you believe you have found a vulnerability, we want to hear from you — and we will respond.

Reporting a vulnerability

Preferred — encrypted, no key exchange needed on your side: open a private report through GitHub, via the repository's Security → "Report a vulnerability" (GitHub Private Vulnerability Reporting). GitHub encrypts the report in transit and at rest.

Alternative: email the maintainer at ruihe93@gmail.com. For end-to-end encrypted email, request our current age public key (or a PGP key) at that address first and we will reply with the recipient before you send any details — or simply use the GitHub private channel above, which is already encrypted.

Please include the affected version or commit, a description, and a minimal reproduction if you have one. Please do not open a public issue for a suspected vulnerability before it has been addressed.

Our commitment (acknowledgement SLA)

  • Acknowledgement: we will acknowledge your report within 72 hours.
  • Triage: an initial assessment (severity and whether it is in scope) within 7 days.
  • Fix and disclosure: we aim to ship a fix and coordinate disclosure within 90 days, keeping you updated along the way. Credit is offered to reporters who want it.

Scope

The governance kernel — infra/ (govd, the executor channel, the verifiers) and the skillChip/ registry — is in scope. The enforcement surfaces (govd.py, executor.py, oversight.py, the *verify.py cores) are the highest-value targets. Findings that bypass a governance boundary are especially welcome: a refusal that should have fired but did not, a value or secret crossing the agent↔govd boundary, or a step running unblessed.

The delegated execution boundary (SV-3) is also in scope: exod (infra/exec/exod.py), the closureverify.py / grantverify.py / exodverify.py cores, and the confined sandbox (infra/exec/sandbox.py). High-value findings there include a forged exod signature accepted as authoritative status, a grant-nonce replay that goes undetected, a step running as root despite exec-never-root, or a sandbox confinement escape.

There aren't any published security advisories