Aegis Guardian
A free, bilingual (English / Español) AI scam shield for everyday people and their families.
Un escudo gratuito y bilingüe contra estafas, hecho para las familias.
Paste a suspicious text, email, or phone number, or upload a screenshot. Aegis tells you if it's likely a scam, shows you the red flags, and tells you what to do next, in plain English or Spanish.
Aegis is a consent-first personal safety and anti-scam agent, built so it can never become the surveillance tool it's meant to prevent. It's part of the Guardian security portfolio, alongside VPN Guardian and Wi-Fi Guardian. Like its siblings, it's open source and auditable, so you can verify there's no hidden backdoor.
Live: try the scam checker in English or Spanish. AI layer powered by Meta's Muse Spark.
Why this exists
Scams, not hacks, are where people lose the most money:
- $20.9 billion in cybercrime losses were reported to the FBI in 2025, a record, from over 1 million complaints.
- 85% of those losses came from cyber-enabled fraud: scams, not break-ins.
- People 60 and over lost $7.75 billion, up 59% in one year.
- Phishing and spoofing was the most reported crime type.
- AI-related complaints were counted for the first time: 22,364 complaints and $893 million lost.
Source: FBI IC3 2025 Internet Crime Report (April 2026).
Data breaches make it worse. Leaked personal info lets scammers write texts that mention your real name, bank, or recent order. Existing scam checkers mostly give a yes/no answer, in English, behind a paid antivirus plan.
Aegis is different:
- Bilingual by default. Spanish-speaking families are heavily targeted and underserved.
- Explains the why. It teaches people the red flags so they get better at spotting scams on their own.
- Family-first. It turns a scam into a shareable warning card you can send to your group chat.
- Free and open source. No upsell, ever.
Most "family tracking" apps quietly cross the line into surveillance. Aegis is the opposite bet: a safety tool people install because it protects them, built transparently enough that they can trust it.
How the scam scanner works
Aegis uses two layers. A fast, local rule engine (already shipped, see app/scam/) catches known patterns with no network at all. The AI layer reads what rules can't: screenshots, new scam scripts, and context.
Screenshot / text / number
│
▼
┌─────────────────┐
│ Rule engine │ local, instant, no network
│ (scam_shield.py)│ known scam patterns + scoring
└────────┬────────┘
│
▼
┌─────────────────┐
│ Muse Spark │ reads the message or screenshot, reasons about
│ (Meta Model API)│ red flags, searches for reports on the sender
└────────┬────────┘
│
▼
┌─────────────────┐
│ Verdict card │ 🔴 Likely scam / 🟡 Suspicious / 🟢 Looks safe
│ │ + red flags found + what to do next (EN/ES)
└────────┬────────┘
│
▼
┌─────────────────┐
│ Muse Image │ optional "Anatomy of this scam" warning card
└─────────────────┘ to share with family
Red flags Aegis checks for
- Fake urgency ("your account will be closed in 24 hours")
- Look-alike or shortened links
- Spoofed senders (banks, USPS, toll agencies, the IRS, Amazon)
- Requests for gift cards, crypto, wire transfers, or payment apps
- Requests to text a different number or move to WhatsApp, Telegram, or another app
- Fake recruiter "task scams" and too-good-to-be-true pay
- "Wrong number" openers and investment offers
- Impersonation of family ("Hi mom, this is my new number" / "Hola mamá, este es mi nuevo número")
- Spanish-language versions of job, urgency, and payment scams
Try it now
In your browser: jeffreyesdavid.github.io/aegis-guardian/app. English and Spanish, runs entirely on your device.
From the terminal:
python3 app/scam/scam_shield.py --file app/scam/samples/scams.txt
Tests: cd app/scam && python3 -m unittest (23 tests) and node app/scam/test_parity.mjs (checks the browser and Python engines agree).
Built with
| Tool | Role in Aegis |
|---|
| Python rule engine | Local, offline first pass (app/scam/) |
| Muse Spark (Meta Model API) | AI scanner: reads text and screenshots, reasons about red flags, web-search grounding |
| Muse Image | Generates shareable scam warning cards |
| SAM 3.1 | Experimental: "X-ray mode" to highlight suspicious parts of a screenshot |
| Muse Code | AI coding agent used to build the project |
Roadmap
- Phase 1 — Scam shield. ✅ Shipped. A local, read-only scanner that flags scam texts and explains why. No network, Python stdlib only.
- Phase 1.5 — AI scam shield (Oct 2026)
- Phase 2 — Safety circle. Opt-in check-ins and trusted contacts, with E2E encryption from day one.
- Phase 3 — Routine sense. On-device anomaly detection and the gentle check-in prompt.
- Later — Private mode. Run the AI layer on open-weight models so sensitive messages never leave your control.
The bigger idea: consent-first safety
A safety agent that protects the people who opt in. It flags scams, lets you share your location with people you choose, and reaches your trusted circle if something's wrong, without ever tracking anyone who didn't agree to it.
The hard line: consent is the only difference between a guardian and stalkerware. Everything here is designed around that.
| Feature | What it does | Data source |
|---|
| Scam shield | Flags scam texts, numbers, and screenshots and explains why | Your messages, shared by you |
| Safety check-ins | Share live location with contacts you pick; auto-share last location if you miss a ping | Your device, opt-in |
| Trusted circle | Name your emergency contacts; the right person is reached first | Your contacts, your choice |
| Routine sense | Notices big breaks from your normal pattern and prompts a gentle "you good?" | On-device only |
What it will never do
- Track someone who didn't opt in
- Tap nearby phones, cameras, Bluetooth, or Wi-Fi to locate a person
- Read a location through other people's devices
- Run any "silent" mode the tracked person can't see and disable
If a feature can't be built with the tracked person's knowledge and consent, it doesn't get built.
Privacy
- Aegis never asks for passwords, bank info, or account numbers.
- The local rule engine runs fully offline.
- The AI scanner sends only the message or screenshot you choose to scan to the Meta Model API. Blur personal details before uploading.
- Scans aren't stored by default.
Security model
A safety app that gets hijacked becomes the worst-case tool. Security is the backbone, not a feature.
- End-to-end encryption. Location and messages are encrypted on-device; the server only ever holds ciphertext it can't read.
- On-device by default. Routine learning and pattern matching run locally. Data that never leaves the phone can't be stolen from a server.
- Hardware-backed keys. Keys live in the secure enclave (iOS Keychain / Android Keystore), never in app files. A copied database is useless without the device.
- No self-escalation. Sensitive actions (add a contact, change who gets alerts, enable sharing) require on-device user confirmation. No remote command can grant itself new access.
- Tamper and takeover alerts. New-device logins, permission changes, and config edits all notify the user immediately.
- Open source and audited. Public code, reproducible builds, documented threat model.
Threat model (who we're defending against)
- A thief with your unlocked phone → hardware keys + confirmation on sensitive actions
- A server breach → E2E encryption; server holds no readable data
- A malicious contact → contacts see only what you shared; no hidden escalation
- An abuser trying to flip on silent tracking → impossible by design; all sharing is visible and user-controlled
Full details: docs/threat-model.md
Architecture outline
┌─────────────────────────────────────────────┐
│ MOBILE APP │
│ (iOS / Android — most logic lives here) │
│ │
│ ┌───────────┐ ┌────────────┐ ┌──────────┐ │
│ │ Scam │ │ Routine │ │ Trusted │ │
│ │ classifier │ │ engine │ │ circle │ │
│ │ (on-device)│ │ (on-device)│ │ manager │ │
│ └───────────┘ └────────────┘ └──────────┘ │
│ │ │ │ │
│ ┌──────────────────────────────────────────┐ │
│ │ Crypto layer (secure enclave keys, E2E) │ │
│ └──────────────────────────────────────────┘ │
└──────────────────────┬────────────────────────┘
│ ciphertext only
▼
┌─────────────────────────────────────────────┐
│ RELAY SERVER │
│ Stores & forwards encrypted blobs. │
│ Cannot read location, messages, or contacts. │
│ Handles: delivery, push triggers, auth. │
└─────────────────────────────────────────────┘
Components
Mobile app (the core)
scam/ — scam scanner: local rules today, AI layer next; reporting pipelineroutine/ — on-device pattern engine; learns normal behavior; flags anomalies; never leaves the devicecircle/ — trusted-contact management; invites, consent records, revocationcrypto/ — key generation in the secure enclave; E2E encrypt/decrypt; key rotationcheckin/ — the ping → wait-window → last-location flow, with cancel
Relay server (deliberately dumb)
- Stores and forwards encrypted blobs only
- Handles auth, push-notification triggers, and delivery
- Knows that two users are connected, never what they share
- No plaintext location, messages, or contact graph ever touches it
Emergency flow
Contact pings you
→ you don't respond within the set window
→ your last location is shared with your chosen circle
→ you can cancel at any point
Repo structure
aegis-guardian/
├── README.md
├── docs/ # landing page + threat model
├── app/
│ ├── scam/ # scam shield (shipped) + AI layer (in progress)
│ ├── routine/
│ ├── circle/
│ ├── crypto/
│ └── checkin/
├── relay/ # store-and-forward server
└── LICENSE
License
MIT
