Ward authenticates login, sudo and polkit by face. This document states what it does and does
not protect against, and how to report a vulnerability. Please read the threat model before filing a
security issue — several "bypasses" are known and documented limits, not bugs.
Do not open a public issue for a security bug. Use GitHub's private vulnerability reporting (the Report a vulnerability button under the repository's Security tab), which opens a private advisory visible only to the maintainers.
Please include: the ward version (ward --version), your distro and camera module, and the
steps to reproduce. If the report involves a presentation attack (photo, screen, mask), a short
video or the recorded frames (ward lab --record-frames) helps enormously.
We aim to acknowledge within a few days and to fix confirmed, in-scope issues before any public disclosure. There is no bug-bounty program.
Every authentication failure — a dark or busy camera, no match, a suspected spoof, a timeout, a crash in the engine, a stuck daemon — must fall through to the password prompt. A bug that locks a user out is treated as the most serious class of defect here, ahead of any false-accept concern. The PAM module enforces its own deadline and never returns "authenticated" on an error path. If you find a way to make ward deny a login rather than fall back to the password, that is in scope and important.
Ward is strong convenience with a password fallback, in the same class as Windows Hello on the same 2D camera — not the class of Face ID (which has a depth sensor) or a hardware security key.
In scope (report these):
- A path where authentication does not fail open (a lockout, or auth granted on an error).
- Privilege escalation: anything that runs attacker-controlled code as root through the PAM module,
the
ward-authhelper, the resident daemon (ward.service), thepkexecpaths used by the app, orinstall.sh/packaging. - The resident daemon's socket accepting a request from a non-root process, or one attempt's state leaking into another.
- Secret handling: the sealed keyring credential,
PAM_AUTHTOK, or the model store (0600) being exposed to a non-root user. - A presentation attack that also defeats the full stack — see below for what each layer does.
Out of scope (known, documented limits of a 2D sensor):
- Identical twins and strong look-alikes. No 2D face system, Windows Hello included, stops these.
Mitigate with
require_rgb_match, a tightcertainty, and measuring the relative inward lab. - A high-quality printed IR photograph: a 2D system cannot fully distinguish it. The anti-spoof layers raise the bar but do not claim to defeat a determined, well-resourced attacker.
- Coercion (being forced to look at the camera). Face unlock cannot defend against this; use the password/lock-out, and keep the password fallback strong.
- A false-accept rate (FAR). Ward does not yet claim a calibrated FAR — that needs a lab study with other subjects (roadmap). Do not treat it as certified.
A single trick that fools one layer is usually not a system bypass:
- Recogniser decides who, not whether it's real. Fooling it needs a look-alike or a photo of the enrolled person — and that photo still has to pass liveness.
- Cross-spectrum (IR) is the real anti-replay: a phone/monitor emits ~no infrared, so a replayed face is absent from the IR frame and the attempt FAILs. This is why ward uses an IR camera.
- Anti-spoof CNN catches the physical artifacts of a camera filming a screen or print (moiré,
reflection, lack of depth). It does not flag a digitally re-encoded image — verified by
ward benchmark redteam— so it is a complement to the IR layer, not a replacement. - A suspected spoof (a liveness FAIL from cross-spectrum / shadow / CNN) ends the attempt. An
optional spoof lock-out (off by default, since a lockout only creates denial-of-service while
the password always works) can pause face unlock after repeated spoofs —
sudo ward tune security.lockout_failures Nturns it on.
You can measure these yourself: ward benchmark lfw (impostor rate, recogniser),
ward benchmark spoof (APCER/BPCER, CNN), ward benchmark ir (cross-spectrum PAD), and
ward benchmark redteam (digital attacks on your own frames).
Ward is pre-1.0; only the latest released version receives security fixes.