TARCHIA: Shadow AI Guardian

💡 Inspiration
This whole enterprise AI boom kind of looks glamorous, but beneath it there is this massive, hidden mess called Shadow AI… and yeah it’s a crisis. Companies are throwing millions 💰 at “secure” AI tools, and yet employees still, kinda quietly, use free consumer web options to move faster. The catch is they’re uploading all kinds of sensitive stuff 🔐, like corporate source code, customer database queries 📊, and even private API keys 🔑 to public servers. So, data leaks 💥 and compliance failures start stacking up fast. We built TARCHIA so CEOs and security teams get direct, no-blink visibility and real-time control over this unhinged, Wild-West AI terrain.

🛡️ What it does
TARCHIA is an autonomous Enterprise AI Security Agent that monitors and protects every employee-to-AI interaction in real time. It does a bunch of things, not just one:

👁️ Intercepts & Analyzes: It automatically checks outgoing data payloads being sent to outside AI platforms.
📊 Calculates Risk Scores: It assigns threat levels from 0 to 100, based on whether a tool is approved and how sensitive the data is.
🛑 Enforces Guardrails: It blocks high-impact leaks instantly (for example PCI-DSS related financial structures, or Base64 encoded keys) before any data leaves the corporate perimeter.
🧠 Tracks Behavioral Anomalies: It spots insider threat vibes, odd usage volume, and those sneaky tool-hopping evasion tactics.

⚙️ How we built it
We powered TARCHIA through the Google Cloud ecosystem ☁️:

Vertex AI Agent Builder: The core multi-step reasoning route, plus orchestration layers.
Google Gemini Enterprise: The deep semantic brain 🧠 that handles complicated analysis, decodes obfuscated payloads, and produces forensic logs.
Google BigQuery: A centralized telemetry data home 📈 to record high-volume traffic patterns and help predictive fraud detection run better.

🔥 Challenges we ran into
A security agent has to be insanely precise—false positives 🚨 can slow employees down, false negatives 💀 can basically wreck security. We managed it by running tests across 9 tough, real-world situations ✅: ChatGPT with credit cards, Claude with API keys, approved tools plus passwords, Base64 obfuscation, tool-hopping patterns, volume anomalies, insider threat attempts, dark web access attempts, and a few more cases we really did not want to mess up.

🏆 Accomplishments that we're proud of
We built a fully operational security control plane inside Google AI Agent Builder. It handles messy data decoding, behavioral analysis, and real-time compliance enforcement. Basically we showed that AI intelligence 🤖 can be used to police and secure AI deployment across a global workforce, without relying on vibes or manual cleanup. 🌍

📚 What we learned

• Security cannot stop at URL filtering—content inspection has to keep going continuously even on approved tools 🔍
  
• Behavioral anomalies (volume, timing, patterns) matter as much as tool approval status 📊
  
• AI can be trained to catch evasion tricks (obfuscation, tool-hopping) in real time ⚡
  

🚀 What's next for TARCHIA: Shadow AI Guardian
🏢 Enterprise pilot deployments with Fortune 500 companies
📊 MongoDB Atlas integration for high-volume data logging
👑 A custom dashboard for CEO/CISO command centers
🌐 API gateway integration for intercepting real network traffic
🧠 Predictive threat modeling using historical decision logs

Built With

• agent-builder
• google-bigquery
• google-cloud
• google-gemini
• vertex-ai