Phishing Detection On-Device: How It Works and Why It Matters
Phishing remains the most common cyberattack vector, accounting for over 90% of data breaches according to Verizon's 2024 DBIR. Here is how on-device AI can help you reason about phishing emails before you click — without sending your messages to a server.
| Stage | What happens | What does NOT happen |
|---|---|---|
| 1. Intake | Email text arrives via Share Extension or paste | Email is not uploaded anywhere |
| 2. PII masking | Names, emails, phone numbers, addresses are tokenized before the model sees the text | PII is not logged, not sent off-device |
| 3. FoundationModels inference | The on-device model produces a summary, risk flags, deadlines, and a draft reply | No network call; no cloud LLM; no vendor API |
| 4. Structured result | Summary, risk flags (with verbatim quotes), deadlines, and a cautious suggested reply are rendered on screen | No behavioral profile is built; no analytics emitted |
| 5. Optional history | The analysis is stored locally (configurable retention) | History never syncs to any server |
Walk into any cybersecurity conference and you will hear about AI-powered threat detection, behavioral analytics, and cloud-scale phishing prevention. These are impressive technologies, but they share a fundamental limitation: they require your email content to be processed on someone else's servers. For individual users — not enterprises with security operations centers — this creates an impossible tradeoff: expose your messages to a third party to protect against phishing, or keep your messages private and risk falling for a scam.
On-device analysis reframes this tradeoff. Modern iPhones running iOS 26 with Apple Intelligence can run Apple Foundation Models locally — large enough to reason about the language of a suspicious email (urgency cues, impersonation patterns, requests for action), without requiring any email content to leave the device.
What Veilsort Actually Analyzes
Veilsort is not a header, SPF/DKIM, URL-destination, or attachment analyzer. v1.0 is text-only. What it does is take the body of a suspicious email — shared from Mail.app or pasted directly — and produce a structured analysis: a plain-English summary of what the sender is really asking for, risk flags with a domain (legal, financial, employment, privacy, reputation, or safety) and a severity of low, medium, or high, explicit or inferred deadlines with verbatim source quotes, and a cautious, non-committal suggested reply. Each risk flag includes a rationale and a verbatim quote from the email, so you can verify the model's reasoning against the original text.
Veilsort does not inspect email headers, validate SPF/DKIM, follow redirect chains to inspect link destinations, analyze attachments, learn from your inbox over time, or build a personalized sender baseline. Each analysis is independent and text-only. Anything that would require ongoing access to your inbox or persistent data about your communications is out of scope — not because it is not useful, but because it is incompatible with the zero-network, zero-account architecture that makes Veilsort trustworthy for sensitive emails.
The Pipeline, End to End
When you share an email into Veilsort, it runs a fixed on-device pipeline. First, PIIMasker strips personal identifiers — names, phone numbers, addresses, URLs, email addresses, and similar — using NSDataDetector, NLTagger, and a set of regex layers, replacing them with token placeholders before the text reaches the model. Second, if the email is long, TextChunker splits it into chunks for a map-reduce pass. Third, Apple Foundation Models produce the structured EmailAnalysis — summary, risks, deadlines, and reply draft. The result is rendered into four cards on screen. None of these steps makes a network request; the only thing that ever leaves the device is the Apple system-level MetricKit diagnostics, which is not initiated by Veilsort and contains no email content.
Why On-Device Matters for Phishing
Phishing emails often contain the very information attackers are trying to steal — your bank's name, your employer's domain, your colleagues' email addresses. Sending these details to a cloud scanner creates a new target. NIST's Phish Scale framework recommends that phishing detection tools minimize data exposure as a core design principle. On-device analysis is not just a privacy preference — it is the only architecture that aligns with NIST's guidance on protecting sensitive content during security analysis.