Two layers of protection. Here’s exactly what each one checks.
A free engine that runs instantly on the device with no network call and no AI, and an optional AI deep check for the hard and the brand-new cases.
Where it works
Gmail (incl. Workspace & custom domains) · Outlook (web & add-in) · WhatsApp · Messenger · Facebook · any website
Install
Chrome & Edge extensions · Outlook add-in · Windows & Mac
Languages
English · Icelandic · Swedish · Finnish
Verdicts
Likely safe · Suspicious · Likely fraud
Basic check
100% on device, no network, no AI, nothing sent
AI deep check
Processed in the EEA · used once · never stored, never sold
Updates
Engine & blocklists refresh without a store release
Basic: on your device
Every signal the local engine checks, with no network call.Domain & sender spoofing
- Brand impersonation: claims a brand in the sender or subject but is not from its domain
- Look-alike domains: 1–2 character edits from a real brand (paypai.com)
- Homoglyph swaps: digits for letters (paypa1.com)
- Punycode / IDN look-alike domains (xn--…)
- Brand hidden in another domain: secure-paypal-login.com
- “@”-in-URL trick: real host hidden after an @
- Free-email account posing as a company
- Malformed / machine-generated sender strings
Email authentication
- SPF, DKIM & DMARC result checks from the headers
- From-domain alignment: a brand From not actually sent or signed by that domain
- Reply-To mismatch: replies routed to a different domain
- Reply-redirect: “reply to this private address instead”
Links
- High-risk top-level domains
- URL shorteners hiding the real destination
- Raw IP-address links
- Redirect / link-laundering through tracker domains
- Link text ≠ destination: shows one site, goes to another
- Brand-new / unlisted domains behind login & pay requests
Social-engineering tactics
- Urgency + pressure to act now
- Credential requests: password, PIN, card number or 2FA code
- Account-verification pressure
- Secrecy requests: “keep this confidential”
- Fake “held / failed messages”: click to release
- Copyright / legal-threat extortion
- Generic greetings (“Dear customer”)
Money & payment fraud
- Advance-fee & business-email-compromise
- CEO / executive-impersonation payment asks
- Invoice & bank-detail change: a new IBAN is a payment redirect
- Gift-card & crypto demands, incl. raw wallet addresses
- Investment / “pig-butchering” lures
- Fake prize / lottery (“you won €…”)
- Family impersonation: “hi mum, new number”
- Fake job offers & fake quote-request openers
Attachments & obfuscation
- Executable & script files (.exe, .js, .scr…)
- Macro-enabled Office docs & macro-bait wording
- Double-extension files (“invoice.pdf.exe”) & HTML attachments
- Zero-width / invisible characters stripped so hidden scam words are caught
- Diacritic-insensitive matching (Icelandic & Swedish spellings)
- Known-good domain allowlist (top sites) to avoid false alarms
Premium: AI deep check
What the optional, user-triggered AI layer adds on top.- Claude AI reasoning: a plain-language explanation of why, in any language, that catches new and unseen scams by understanding intent, not just matching patterns.
- “Is this site safe?”: checks the website you are on for fake sign-ins, brand spoofing and look-alike pages.
- Domain-age check: looks up the sender and link domains’ real registration dates (registry RDAP + certificate transparency) and feeds them into the AI’s reasoning, so a recently-registered domain raises the verdict instead of being mistaken for an established one.
- Scam-category classification (phishing, smishing, investment, bank & government impersonation, delivery, and more) with a matched 60-second lesson.
- EEA processing: runs on Amazon Bedrock in the EU, per-partner region (EU / US / Australia); the item is used once to produce the verdict, then discarded, never stored, never sold.
How a verdict is reached. Signals are weighted into a score and mapped to one of three plain labels (Likely safe, Suspicious, Likely fraud) with the specific red flags shown. Confidence is scored internally and capped (never presented as certainty): SafeAction is an assistant that keeps people alert, not a guarantee, and a “likely safe” result still advises caution.