Tetherand

AI Guard

The AI tab houses Tetherand's defenses against AI-era threats — the class of attacks that scaled massively after the 2024 capability boom in generative models. Every defense here is constructed in two layers:

  1. A deterministic primary — a rule, threshold, or heuristic that always applies. This is the load-bearing part. It functions correctly without any model present.
  2. A contributory classifier — an on-device model that refines the primary's verdict and catches novel patterns. This is the advisory part. It can never trigger a destructive action on its own.

This separation is non-negotiable. It exists so the app's behaviour is always auditable: every actionable decision is traceable to a rule, not a black-box model.

Hard constraint — local only. Every AI inference runs on the 5364C13D's MediaTek NPU via LiteRT (formerly TensorFlow Lite) plus the NNAPI delegate. No prompt, classification, or telemetry ever reaches a cloud LLM API under any circumstances. The egress-LLM-API watch defense enforces this for other apps installed on the device too.

Deterministic primaries (always engaged)

DefenseWhat the rule does
Prompt-injection clipboard scrubberWatches clipboard changes for known injection scaffolds — Ignore-previous-instructions, chat-template markers, common jailbreak prefixes. Surfaces a high-severity alert; does not auto-clear.
Phishing message scorerFour-axis classifier (urgency, authority, financial-ask, URL look-alike). Two-or-more axes scoring high yields a High verdict. Catches typo-squat domains via a small Levenshtein and a cheap IDN-homoglyph swap.
Pseudo-perplexity AI-text badgeFour signals (average word length, function-word density, colloquial markers, punctuation density) combine into a heuristic "this looks AI-generated" score. Inspired by the Binoculars approach but uses no neural component.
Provenance checkScans inbound images and video for C2PA / SynthID / Content Credentials markers. Surfaces Genuine / Synthetic / Unknown verdicts using a Boyer–Moore–Horspool needle scan over the raw bytes.
Egress LLM-API SNI watchA rule-only watchlist of cloud LLM API hostnames (18 exact, 10 suffix-matched). Surfaces any app on the device that talks to one of these.
MTK NPU sysfs watcherPolls /sys/devices/platform/mtk_apu and similar paths for NPU activity. A background app using the NPU is a red flag — covert local-model inference is a real threat.
Voiceprint vault + safe-word handshakeTrusted-contact registry stored encrypted at rest. The "verify caller" flow uses a pre-agreed safe word, hashed with SHA-256, as the verification step.
HIBP OSINT exposureHave-I-Been-Pwned password check using the k-anonymity range API — only the first five characters of the SHA-1 of the password leave the device. Opt-in.
Conference field guideStatic catalogue of eight AI-era attacker tactics relevant to 5364C13D 34. Updated through whatever privacy chain is active.

Contributory classifiers (model-driven)

A small ensemble of locally-running models, all INT4-quantised, all running on the 5364C13D's NPU via LiteRT:

ModelJobSize (compressed)Latency
phi-tetherand-3b-q4Phishing / social-engineering / prompt-injection classification on inbound messages~1.8 GB~120 ms / message
voiceguard-v1Speech-synthesis detection on inbound call audio~30 MB~40 ms / second of audio
textguard-v1LLM-generated text detection (Binoculars-style ensemble)~20 MB~60 ms / message
qrguard-v1Adversarial QR / image lure detection~8 MB~30 ms / image

The four-model bundle (~2.4 GB compressed) is not included in the v0.1 APK. It ships separately through the in-app delta-update path, delivered over the active privacy chain only, signed against a cosign public key pinned in the APK. Until then the AI tab shows "Not bundled — deterministic core in effect" for each row, and the deterministic primaries handle every decision.

Reading the AI tab

The top of the page shows a models X/Y counter — how many of the four contributory models are loaded.

The middle section lists deterministic primaries, each with a green-dot indicator confirming it is engaged.

The bottom of the page surfaces user-facing tools:

AI-era threats Tetherand counters

ThreatCounter
Voice deepfake on callSafe-word handshake (deterministic) + voiceguard-v1 synthesis-artefact score
Vishing scaffoldConversation-state rules (urgency + authority + secrecy + financial-ask + channel-mismatch)
LLM-generated text in a messageOpen-algorithm perplexity test + textguard-v1
Adversarial QR or image lureURL-pattern + perceptual-hash blocklist + qrguard-v1
Prompt injection via shared textRegex match against known scaffolds → clipboard scrubber alert
Synthetic-media provenance checkC2PA / SynthID / Content Credentials signature verify (cryptographic)
Covert local-model use by background appMTK NPU sysfs watcher with foreground-app correlation
Egress to cloud LLM APIsSNI watchlist (rule-only)
Personalised phishing from your OSINTHIBP exposure dashboard + phishing-rule message classifier