惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

GbyAI
GbyAI
Martin Fowler
Martin Fowler
I
InfoQ
腾讯CDC
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
爱范儿
爱范儿
Microsoft Security Blog
Microsoft Security Blog
Google DeepMind News
Google DeepMind News
D
DataBreaches.Net
云风的 BLOG
云风的 BLOG
F
Fortinet All Blogs
N
Netflix TechBlog - Medium
博客园 - 聂微东
Microsoft Azure Blog
Microsoft Azure Blog
D
Docker
博客园 - 三生石上(FineUI控件)
Y
Y Combinator Blog
博客园 - Franky
Engineering at Meta
Engineering at Meta
B
Blog
罗磊的独立博客
Apple Machine Learning Research
Apple Machine Learning Research
Jina AI
Jina AI
V
Visual Studio Blog

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
I Built a Web3 Threat Intelligence Oracle Powered by Gemma 4
James Kabing · 2026-05-18 · via DEV Community

This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

OCTIO (On-Chain Threat Intelligence Oracle) is a system that monitors live phishing and malware feeds, runs each threat indicator through Gemma 4 for analysis and classification, and stores verified intelligence in a format that blockchain protocols can query before executing sensitive operations.
The problem it solves is real. Around 80% of funds stolen from Web3 projects come not from smart contract exploits but from attacks on surrounding Web2 infrastructure such as phishing campaigns, DNS hijacking, frontend supply chain compromise, and cloud misconfiguration. The Ronin Network hack ($625 million) started with a phishing email. The Ledger Connect Kit attack ($600,000) came from a poisoned npm package. The Curve Finance DNS hijack redirected legitimate users to a malicious server for hours.
In every case, observable signals existed before significant losses occurred. OCTIO is the infrastructure to collect, analyse, and distribute those signals with Gemma 4 at the core of the intelligence layer.

Demo

DemoGitHub: github.com/James-Kabingu/octio

Running the full system takes five commands:
python3 monitor.py # fetch and classify live phishing URLs
python3 registry.py # submit validated indicators to registry
python3 oracle.py # run protocol query interface
python3 correlation.py # correlate against documented incidents
python3 dashboard.py # display full system dashboard

OCTIO Dashboard

Sample oracle output:
[BLOCK ] [CRITICAL ] http://www.dpdlocoqu.cyou/com
[PROCEED] [SAFE ] https://uniswap.org
[CAUTION] [SUSPICIOUS] http://metamask-security-alert.com/connect
[PROCEED] [SAFE ] https://aave.com
[BLOCK ] [CRITICAL ] http://instagram.com.universal-api.org/

Code

GitHub: github.com/James-Kabingu/octio
The repository includes:

monitor.py: Live phishing feed monitoring with Gemma 4 threat classification.
registry.py: On-chain registry simulation with keccak256 hash storage.
oracle.py: DeFi protocol query interface with Gemma 4 risk assessment.
correlation.py: Incident correlation against documented real-world hacks.
dashboard.py: Terminal dashboard for live threat visibility
contracts/ThreatRegistry.sol. Solidity contract for Sepolia testnet deployment.

The smart contract implements the full indicator submission, validation, and query interface described in the system architecture. Target domains are stored as keccak256 hashes rather than plaintex, reducing storage costs by 60 to 90% and preventing the registry from being used as a phishing directory.

How I Used Gemma 4

I chose the Gemma 3 27B Instruct model (google/gemma-3-27b-it) via OpenRouter. The 27B model has the reasoning depth needed to make nuanced security judgements, not just pattern matching.
Gemma 4 powers three distinct functions inside OCTIO:

  1. Threat Classification (monitor.py): OCTIO pulls live URLs from OpenPhish, a real phishing feed updated continuously. Each URL is sent to Gemma 4 with a structured prompt asking it to classify the threat, assess severity, identify the impersonation target, and explain its reasoning. Given the URL http://www.dpdlocoqu.cyou/com, Gemma 4 returned:

{
"is_threat": true,
"threat_type": "PHISHING",
"severity": "HIGH",
"target": "DPD (Delivery Service)",
"reasoning": "The URL uses a misspelling of DPD and a suspicious domain extension to impersonate the legitimate delivery service, likely to steal credentials potentially leading to crypto access."
}

This is not keyword matching. Gemma 4 understood the typosquatting pattern, identified the impersonated brand, and connected it to the crypto threat context from the URL alone.

  1. Risk Assessment (oracle.py):
    When a protocol queries the oracle, Gemma 4 produces a final risk assessment combining registry data with its own domain pattern analysis. For http://metamask-security-alert.com/connect, which was not in the registry, Gemma 4 returned SUSPICIOUS/CAUTION based on the domain pattern alone. A rule-based system would have returned CLEAN. Gemma 4 understood the social engineering intent from the URL structure.

  2. Incident Correlation (correlation.py):
    Gemma 4 analyses the current indicator set against a database of documented real-world Web3 hacks. Given five live phishing URLs, it correctly correlated them with the Ronin Network hack ($625M), Curve Finance DNS hijack ($570K), and MyEtherWallet BGP attack ($17M), and produced actionable recommendations for DeFi protocols.

Why the 27B model:
The 27B model was necessary for this task. Smaller models returned surface-level classifications without the reasoning depth needed to distinguish between a legitimate security alert and a phishing site designed to look like one. What Gemma 4 unlocked was a system that gets smarter as the threat landscape evolves, without retraining or updating rule sets.

Limitations and Next Steps

This is a working prototype, not a production system. The current limitations are worth being clear about:
The on-chain registry is simulated in Python. The Solidity contract (ThreatRegistry.sol) is written and ready but not yet deployed to Sepolia testnet. A live deployment requires testnet ETH and a Foundry deployment script.
The monitoring layer currently pulls from OpenPhish only. The full architecture specifies Certstream for certificate transparency monitoring, PassiveDNS for DNS change detection, SecurityTrails, and npm audit feeds. These are the next integrations.
The prototype auto-validates every indicator Gemma 4 flags. The whitepaper describes a multi-party validation mechanism where registered researchers vote on submitted indicators before they reach the active registry. That governance layer is not yet implemented.
Gemma 4 can produce false positives. The model occasionally flags legitimate domains as suspicious. Human review before on-chain submission is essential in a production deployment.
The oracle interface simulates Chainlink integration. A real deployment requires a running Chainlink node and a registered External Adapter.
Despite these limitations, the core thesis is demonstrated: Gemma 4 can perform nuanced threat classification, catch suspicious domains not yet in any registry, and correlate current indicators against historical incidents, all in real time. The prototype shows the system works. The remaining work is engineering, not concept validation.