• 

  Feature 1. Inbox Shield: Detects intent-based hijacking in emails using Gemini 3.

• 

  Feature 2. Repo Sentinel: Audits code for "context-aware" leaks and logic flaws.

• 

  Audit logs

• 

  Details of contextual risks & remediation

🛡️ Bastion AI Security Gateway

Full-Spectrum Semantic Immunity for the Agentic Era

---

� About the Project

🛡️ Inspiration

As AI transitions into Autonomous Agents, they face a dual crisis: Infiltration (hijacking via indirect injections) and Exfiltration (leaking system prompts). Research shows agent vulnerability rates $V_{rate} > 26\%$, yet traditional tools show 0% detection because they cannot parse semantic intent. I built Bastion to give agents a "Semantic Immune System."

🚀 What it does

Bastion is an AI Security Gateway that intercepts and audits agent data streams:

• Inbox Shield: Detects intent-based hijacking in emails using Gemini 3.
• Repo Sentinel: Audits code for "context-aware" leaks and logic flaws.
• Privacy Layer: Local PII redaction ensuring $\text{Data}_{leak} = 0$.
• Canary Defense: Injects invisible markers to trace and block covert leaks.

🏗️ How I built it

I utilized React 19 and Vite for a high-performance dashboard. The core engine is Gemini 3 (Flash), integrated via the Google Generative AI SDK, paired with the Gmail and GitHub REST APIs for live data auditing.

🚧 Challenges I ran into

The primary challenge was the non-deterministic nature of AI. I solved this by implementing a Heuristic-AI Hybrid Layer, where Gemini provides reasoning, but a deterministic auditor (Canary check) makes the final security decision. I also built a custom retry-service to handle API rate limits.

🧠 What I learned

I learned that in the agentic era, Context is the new Firewall. Security is no longer about blocking "bad words" but understanding hierarchical instruction priority through Gemini 3's Long Context window.

---

🏗️ Architecture

graph LR
    A[Gmail/GitHub] --> B(Bastion Shield/Sentinel)
    B --> C{Gemini 3 Engine}
    C -->|Threat| D[Quarantine/Patch]
    C -->|Safe| E[Agent/User]
    F[Canary Tracker] -.->|Audit| D

�️ Tech Stack

• AI Core: Gemini 3 API, Google Generative AI SDK
• Frontend: React 19, Vite, Tailwind CSS, Lucide
• Integration: Gmail API, GitHub REST API
• Visualization: Recharts

💡 Key Features

1. Inbox Shield (Anti-Injection)

Uses Gemini 3's Long Context to analyze multi-step semantic hijacking attempts in email streams. Automatically flags "System Overrides" that appear legitimate to traditional filters.

2. Repo Sentinel (Context-Aware Audit)

Beyond regex: understands code intent to find leaked business logic, system prompts, and "Digital DNA" before they hit production.

3. Canary Defense

Injects invisible "tripwire" tokens into prompts. If the AI output contains these tokens, a breach is confirmed with zero false positives.

4. Zero-Leak Privacy

All PII (Emails, Keys, CCs) is redacted locally on the client side before any data is sent to the AI engine for analysis.

---

🛠️ Setup

1. git clone [repo] && npm install
2. Add VITE_GEMINI_API_KEY, VITE_GOOGLE_CLIENT_ID to .env.local.
3. npm run dev

Built With

• github-rest-api
• gmail-api
• google-oauth2.-libraries:-recharts
• typescript