FFai is a media and inference toolkit written in Rust. Several of its crates parse untrusted input — audio, images, documents, model files — so we treat security reports as first-class work.

Please do not open a public issue for a security problem.

Report privately through GitHub's private vulnerability reporting:

1. Go to https://github.com/Remade-With-Rust/FFAI/security/advisories/new
2. Describe the issue, the affected crate and version, and how to reproduce it.
3. Include a proof-of-concept input if you have one — a crashing file is worth a thousand words, and we will add it to the regression corpus once fixed.

If GitHub reporting is unavailable to you, open a public issue containing only the words "security report, requesting private channel" and nothing else, and a maintainer will open a private advisory to continue in.

Some FFai crates process personal data — ffai-mercury derives voiceprints, which are special-category biometric data under GDPR Art 9 (see crates/ffai-mercury/docs/data-inventory.md). If your report describes a defect that could expose personal data, say so explicitly in the first line. That starts a 72-hour clock under GDPR Art 33 for any deployer acting as a controller, and we will notify affected integrators within that window rather than waiting for a fix.

We follow coordinated disclosure. We will agree a disclosure date with you, and we will credit you in the advisory unless you ask us not to.

If a report is already being exploited in the wild, tell us — the 90-day clock is not a shield and we will move immediately.

In scope — anything that lets untrusted input cause memory unsafety, a panic that becomes a denial of service, incorrect output presented as correct, or disclosure of data the caller did not intend to expose. Parsers, decoders, and unsafe blocks are the highest-value targets: ffai-mercury (audio, model files), ffai-carmenta (documents, images), ffai-diana, ffai-media.

Out of scope — vulnerabilities in dependencies with no FFai-specific exposure (report those upstream, and tell us so we can pin or patch), issues requiring a compromised local machine or write access to the model cache by design (see each crate's threat model), and results from automated scanners without a demonstrated impact.

FFai is pre-1.0. Security fixes land on the latest published minor of each crate. There is no long-term-support branch yet; when a crate reaches 1.0.0 under the hardening process (docs/plans/use-protection-please.md per crate), this table gains a support window.

Each crate carries a hardening audit against a 41-gate standard plus 14 compliance controls, tracked in crates/<name>/docs/plans/use-protection-please.md and summarised at the bottom of each crate's README. A crate does not ship 1.0.0 until its v1.0.0-blocking gates pass.