CYBERGUARD AI
One platform that catches scams anywhere: phishing texts, fake links, deepfakes, malware and even live phone calls, with AI that explains why.
- Video of it working
- Problem, approach and result explained
- Public source code
- Expert reviewed· 6 reviews
20%-6%
speech to text
12/12
scam keyword caught
1299
automated tests passing
Walkthrough
Overview
The problem
Pretty much everyone we know has got a “your KYC has expired” SMS, or a call from someone claiming to be from the bank asking for an OTP. Scams don’t come in just one form anymore. It’s a link today, a phone call tomorrow, and sometimes a voice note that sounds exactly like a relative.
What bothered us is that the tools are all separate. One app checks links, another checks files, and nothing helps you while you’re actually on the call. That’s where people lose money, on the phone, under pressure. In a company it’s worse, because one employee falling for a fake “IT support” call can expose everyone.
So we wanted one system that catches a scam wherever it shows up and tells you, in simple words, why it’s dangerous.
The approach
We built CyberGuard AI, and all the AI runs locally on the laptop, without paid APIs.
It has separate checks for phishing messages (including Indian and Hinglish scams), malicious links, impersonation, deepfake images, videos and voices, malware files and logs. Each one gives a reason, not just a score. For impersonation, it first figures out who is being faked (bank, government, a boss) and then checks that specifically.
The part we spent the most time on is live calls. Normal SIM calls can’t be monitored by any app, so we set up our own phone server using Asterisk and free calling apps. Both voices get transcribed live with Whisper. Then our engine reads the conversation for meaning, not keywords: “never share your OTP” is fine, “tell me the OTP you just got” is not. It also follows the attack as it builds up (bank claim, then “your account will be blocked”, then the OTP request), so the risk goes up step by step. The dashboard gets an alert during the call. We chose not to cut the call automatically and leave that decision to a person.
We also made it work for a team. One laptop acts as the server over HTTPS on the local Wi-Fi, other PCs log in from their browsers, and a small agent reports each PC’s risky downloads. Dangerous files get quarantined automatically.
The result
Everything works end to end: scan something and you get a verdict with reasons, saved in history and alerts. Live calls between two phones work through our server, and scam lines get flagged while people are still talking. When we tested the speech recognition on noisy phone audio, wrong words dropped from about 20% to under 1% after our fixes, and it caught all the scam keywords (OTP, UPI, Aadhaar) instead of 8 out of 12. The project has 1,299 automated tests passing.
Reflection
The hardest part wasn’t training models, it was context. Our early versions flagged anything with “OTP” or “bank” in it, which is useless, because real bank warnings use the same words. Getting it to tell a warning apart from a request took most of our effort.
We also learned that real testing finds what unit tests don’t. Actual calls showed us a 10-second freeze, and a bug where the system hung up its own listener. Background noise made the transcription mash sentences together. We only found these by making real calls.
There are still limits. Some deepfake checks give false alarms on real photos. Normal phone numbers would need a telecom SIP connection. The voice models also need a decent amount of RAM. Next, we want to connect a real phone number and check downloads before they’re even saved.
Submitted to CYBERGUARD: Trust Nothing, Explain Everything
- Your idea in one line
CyberGuard is an AI-powered multi-modal detection engine that explains and prevents every threat in simple words with its MITRE ATT&CK technique.
- What will be working by the finale?
1.Phishing detection for emails, SMS and social media messages, using our trained models. 2.URL and website checking with a domain model plus live page checks. 3.Account takeover detection from login patterns. 4.Impersonation detection for fake bank, government and “known contact” messages. 5.Automatic protection running in the background: it watches the Gmail inbox, the Downloads folder and sign-ins, and creates incidents on its own. 6.MITRE ATT&CK mapping on every risky result, with an explanation of what the technique means. 7.Incidents, alerts on Telegram and email, IOC memory that flags repeat threats as “Seen before”, and PDF/CSV reports. 8.Secure login with email OTP, admin-only controls, rate limiting and an activity log. 9.Insider threat model trained for detection of cyberthreat. 10.Deepfake image/video detection.
- Tech stack and hardware you plan to use
1.Languages: Python (backend, ML training and detection), TypeScript (frontend), 2.HTML and CSS (UI styling with Tailwind), SQL (PostgreSQL database) 3.Frontend: React, TypeScript, Vite, Tailwind CSS 4.Backend: Python, FastAPI, SQLAlchemy, Alembic 5.Database: PostgreSQL 6.ML: scikit-learn (TF-IDF + Logistic Regression for text, HistGradientBoosting for URL domains, risk-based scoring for login behaviour. 7.Other: Gemini API for the chat assistant, Gmail IMAP, SMTP for OTP emails, Telegram Bot API for alerts, URLhaus threat feed, PyJWT and bcrypt for auth
- What have you built or tested so far?
1.Most of the system is already working end to end on our laptop. You can sign up with email OTP and scan any detection module .we get a risk score from 0 to 100 with the reasons behind it, the MITRE ATT&CK techniques and what to do next to prevent from any vulnerabilty to be happen. High-risk results become incidents and send a Telegram/email alert. 2.Models trained so far: -URL domain model -Email and SMS phishing model -Account takeover model -Impersonation models -Social media phishing model 3.The automatic watchers have scanned our real Gmail inbox and Downloads folder; our own account has over 130 scans and 100+ threat events so far. We also did a full security audit of our own app, fixed the issues we found, and the backend now passes 270 automated tests.
- Pitch deck (PDF)
- Pitch deck (PDF)PDF
- GitHub repository
- github.com/ashisp9337-cell/Cyberdeck-AI
- What's working now?
CyberGuard AI is a working cyber-safety platform that detects phishing, scams, impersonation, deepfakes and malware,network. Most of the AI runs and give result positively You can scan e-mails, SMS (including Indian scam messages), links, QR codes, files, voice notes, images and videos. It tells you if something is a scam, a fake bank or government message, an AI-cloned voice, a deepfake, or a malicious file, and explains why. Dangerous downloads are moved to quarantine automatically. New for the finale is live voice-phishing detection. Calls go through our own phone server, both sides are transcribed live, and CyberGuard flags OTP requests, fake bank or police callers and AnyDesk tricks while the call is still going. It understands context, so “never share your OTP” isn’t flagged, and it alerts the dashboard without ever cutting the call. It also works for a whole team. One PC acts as the server over secure HTTPS, employees sign in from their own PCs, and the admin sees every user and device live. A browser extension checks websites as you browse. It’s backed by 1,299 passing automated tests. Some deepfake checks can still give 1-2 %false alarms, so in unclear cases they ask for a review instead of raising an alert.
- GitHub repository
- github.com/ashisp9337-cell/CyberGuard-AI
- Deck or report (PDF)
- Deck or report (PDF)PDF
AI tools used
help this to fix the error quickly,and running the test efficiently