Ten capabilities, counted honestly.
Each capability lists every feature with its state, and every attack family it maps to with the status our engineering register gives it. A capability is marked available when at least 70% of the features that work today are fully working; otherwise it is limited. Planned features are listed, not hidden.
- 0
- Capabilities
- 5 limited · 0 not released · 1 coming
- 5
- Features tracked
- 55 work today
- 6
- Attack families
- register 2026-10-01, rev. 2026-10-13
- 0
- Detected end to end
- rule + sensor + test
Every module, its state, its features and its attack families
Feature cells come from the capability's feature list. Family cells come from the coverage register; hover a cell for its name.
| Capability | State | Features | Today | Not rel. | Coming | Attack families |
|---|---|---|---|---|---|---|
MOD.01 Scam protection | Limited | 14 | 3 | 3 | PRT 1N/S 3 | |
MOD.02 Phishing protection | Available | 4 | 2 | 2 | PRT 1N/S 3 | |
MOD.03 QR code protection | Limited | 4 | 0 | 0 | N/S 2 | |
| Limited | 4 | 0 | 2 | OBS 1N/S 2 | ||
| Available | 8 | 2 | 2 | PRT 3OBS 1N/S 1 | ||
MOD.06 Identity protection | Available | 4 | 3 | 1 | DET 1PRT 2N/S 4 | |
| Limited | 3 | 1 | 2 | N/S 2 | ||
MOD.08 Network protection | Limited | 6 | 2 | 3 | PRT 2N/S 4 | |
| Available | 8 | 6 | 1 | PRT 7OBS 1 | ||
MOD.10 AI-adversary defense | Coming | 0 | 0 | 3 | N/S 7 |
"Today" counts features that are available or limited. Families can map to more than one capability (for example vishing sits under scam and deepfake), so the family column does not add up to 127.
What each capability is for
Checks a message, payment request or caller against known scam patterns, on the device.
Checks a link before you open it and warns about lookalike domains.
Decodes a QR code and checks where it really points before you pay or open it.
DNS filtering and your own blocked and allowed lists on the phone; page-level browser protection is not built.
Known-malware scans, automatic protection and fake-app checks on the phone; ransomware warnings on servers.
Password exposure checks, account sign-in safety and hacked-account help today; verified-address monitoring interfaces need a configured provider for results.
A file check and guidance only. No detection engine runs, so Sentinel never says whether media is real or fake.
Wi-Fi and connection awareness on phones; port-scan visibility and firewall control on servers.
A risk score built from real device signals: root, screen lock, encryption, risky apps.
Defence against prompt injection, agent hijacking and poisoned models. Not built yet.
127 attack families. 1 detected end to end.
Our engineering register maps every attack family in our threat taxonomy to the sensor that could see it, the rule that looks for it and the test that proves the rule fires. It records what the code can do, not what a roadmap promises.
- Detected
- 01/127
- Partial
- 27/127
- Observable
- 15/127
- No sensor
- 84/127
How a family is counted as detected
The rule is applied strictly; when in doubt, a row is downgraded. All three gates must hold.
- GATE 1 · RULE
A rule aimed at the family
Code that looks for this attack, with a stated threshold — not a label or an ATT&CK mapping.
- GATE 2 · SENSOR
A real sensor feeds it
Something in the product produces the rule's input outside tests. A rule nothing feeds is "inert" and counts as partial.
- GATE 3 · TEST
A test makes it fire
An automated test exercises the trigger condition. Manual scripts excluded from CI do not count.
- RESULTDetected
Miss any gate: partial, observable or no sensor.
- ×An ATT&CK mapping on its own — it only labels a category.
- ×An event category that nothing produces.
- ×Preventive controls such as authorization checks, rate limits and the desktop WARN ceiling.
- ×Manual live-server scripts that the CI pipeline excludes.
- DET
- Detected. A rule for this family exists, a real sensor can feed it outside tests, and an automated test exercises its trigger.
- PRT
- Partial. A rule exists but is untested, covers only part of the family, or has nothing feeding it.
- OBS
- Observable. Telemetry that could show the family is collected, but no rule looks for it.
- N/S
- No sensor. No sensor in our code can observe this family today.
One pipeline behind every capability
Every detection follows the same path. The evidence and action stages are specified, built and tested in the desktop repository, and not yet shipped.
Collect a fact
A log line, a process, a file change, a device setting. Classified at source; restricted data never leaves the device.
LimitedServer agent, Android 4.50 in production; desktop in repositoryMatch a stated threshold
Six hub correlation rules with registry ids, levels and ATT&CK mapping; three are fed in production. Four desktop rules in the repository.
LimitedHub, productionJudge confidence from evidence
No evidence caps confidence at 0.3; heuristics alone cap at 0.8; recurrence never raises it. Severity and confidence are separate axes.
Not releasedDesktop detection engine (repository)Act within the ceiling
Eight frozen actions from ALLOW to RECOVER. Today a safety gate caps every automatic action at WARN; absence of a policy never grants more.
Not releasedDesktop policy engine (repository)
Confidence is not severity
Confidence asks "are we right?"; severity asks "how bad if we are?". Both are shown and never averaged into each other.
Unknown is not safe
Every failure path in reputation resolves to unknown, and unknown never authorises an action or reads as trusted.
Reversible by default
Quarantine is an encrypted container with an authenticated header; a restore that fails any integrity check is refused.
Which edition carries which capability
A dot means the capability is part of that edition. The capability's own state still applies: a dot on a module marked coming is a plan, not protection.
| Capability | Personal | Safeavio Mobile | Family | Premium | for Small Business | Enterprise XDR | Safeavio ServerGuard |
|---|---|---|---|---|---|---|---|
| Scam protection | · | · | · | ||||
| Phishing protection | · | · | · | ||||
| QR code protection | · | · | · | · | |||
| Browser and web protection | · | · | · | · | |||
| Malware and ransomware protection | · | · | · | ||||
| Identity protection | · | · | |||||
| Deepfake and impersonation | · | · | · | · | · | · | |
| Network protection | · | · | · | ||||
| Device health and posture | · | · | · | ||||
| AI-adversary defense | · | · | · | · | · | · |
What may leave a device, for every capability
Data is classified on the device before it is queued. Anything not explicitly classified is treated as restricted.
| Data | Class | May leave the device? |
|---|---|---|
| File contents | restricted | Never |
| Full file paths | restricted | Never — basename only, if a feature needs it |
| Browsing history | restricted | Never |
| Keystrokes, clipboard, screen | restricted | Never |
| A URL you paste into the link check | restricted | Never — verdict and hostname only |
| Link verdict and hostname | sensitive | With consent |
| Detection category, severity, rule, confidence | security | With consent |
| Device name, OS, CPU, memory, disk | operational | With consent |
"With consent" means cloud reporting is turned on. It is off by default on consumer devices, can be turned off at any time, and the server refuses telemetry from a device with reporting off.
Start with what works today.
The Android app and the Linux server agent are available now.