Security

Gako is built so the server is a non-target: a full server compromise reveals no secret content. This page is the operator-facing summary of that posture. For the conceptual explanation, see the zero-knowledge model.

Pre-release

Gako has not been audited. Treat it as experimental and do not store real secrets in it yet.

What the server learns

The server stores opaque ciphertext, access policy, and signatures. From those it necessarily learns metadata: that objects exist and roughly how large they are (ciphertext is bucketed, not measured exactly), when they are created and change, and the access graph — which identities may read which objects. What it never learns is the plaintext of any secret, because it never holds the decryption keys.

Draft

The precise, line-by-line accounting of what the server stores and what it can infer lives in the Gako design document, data-model specification, and threat-model checks. These will be linked here once the source repository is public.

What Gako does not protect against

Making the server a non-target leaves the endpoints — and the choices operators and users make — squarely in scope. Zero-knowledge does nothing for:

• Compromised client devices. A device compromised while unlocked exposes whatever it can already decrypt; endpoint hygiene is outside Gako's reach.
• Weak or reused master credentials. A phished, guessed, or reused master password defeats the account it protects. Encourage strong, unique passwords, and treat recovery codes as offline secrets.
• Misconfigured access. Policy is enforced exactly and honestly as configured — granting the wrong identity grants the wrong identity. Review grants and machine scopes.

The zero-knowledge model frames the same boundary from the user's side.

Reporting a vulnerability

Report suspected vulnerabilities privately — please do not open a public issue for anything you believe is exploitable. The formal channel is GitHub Security Advisories on the source repository, which opens to the public once that repository does; until then, contact the maintainer directly.

What to expect for a confirmed report:

• acknowledgement within 7 days;
• an assessment and remediation plan within 30 days;
• credit in the release notes, unless you would rather not.

There is no bug-bounty program. Several properties are documented non-guarantees, not bugs, so it is worth reading the zero-knowledge model first: the server sees the access graph, timing, and ciphertext size buckets; revocation cannot recall values already read; and recovery material is, by design, an offline all-powerful artifact.