SentinelAI - Real-Time Chat Security Agent

SentinelAI is a real-time cybersecurity system that analyzes messages using both AI reasoning and deterministic rules to detect phishing, scams, malware, impersonation, and prompt-injection attacks.

Comment

Inspiration

Cyber threats no longer arrive as obvious malware or exploits.
They arrive as messages — phishing emails, scam texts, impersonation attempts, and even prompt injection attacks targeting AI systems.

While building and experimenting with AI agents, we realized a critical gap:

  • Rule-based filters (regex, keywords) are too rigid to catch social engineering.
  • Raw LLMs are powerful, but too unpredictable to be trusted with security decisions.

As AI systems increasingly communicate with users — and with other AI agents — who protects the conversation itself?
That question inspired SentinelAI.

What It Does

SentinelAI is a real-time chat security agent that analyzes messages and enforces clear safety decisions:

  • ALLOW – Safe content
  • FLAG – Suspicious content
  • BLOCK – Dangerous content

Instead of acting like a chatbot, SentinelAI functions as an intelligent firewall for conversations, protecting users and AI systems from phishing, scams, malware, impersonation, and prompt injection attacks.

How We Built It

SentinelAI is built using a hybrid architecture that combines deterministic security logic with AI reasoning:

  1. Deterministic Preprocessing

    • Extracts URLs and detects urgency or authority signals
    • Provides fast, reliable signals for analysis

  2. Gemini 3 Reasoning Engine

    • Gemini 3 is used purely for reasoning, not text generation
    • It evaluates intent, social engineering patterns, and adversarial behavior
    • Returns structured JSON output (threat type, confidence, reasoning)

  3. Risk Scoring Engine

    • Converts Gemini’s output into a numerical risk score (0–100)

  4. Policy Engine

    • Enforces deterministic actions (ALLOW / FLAG / BLOCK)

  5. Memory & Escalation

    • Repeated suspicious behavior automatically escalates risk

  6. Audit Logging

    • Every decision is cryptographically hashed (SHA-256)
    • Stored in append-only logs for transparency and trust

The system is deployed on Google Cloud Run, with a FastAPI backend and a Next.js frontend for low-latency, real-time interaction.

Gemini 3 Integration (Why It Matters)

Gemini 3 is the core innovation behind SentinelAI.

We use Gemini 3 as a decision-making and reasoning engine, not as a conversational chatbot. Its ability to:

  • Understand nuanced intent
  • Detect social engineering
  • Identify prompt injection attempts
  • Produce structured, machine-readable outputs

…allows SentinelAI to make explainable, reliable, and production-safe security decisions.

Without Gemini 3’s advanced reasoning capabilities, SentinelAI would not be possible.

Challenges We Faced

  • Designing a system where AI reasoning is powerful but bounded
  • Preventing hallucinations from influencing security decisions
  • Ensuring explainability and auditability for every action
  • Deploying a full-stack system that is both fast and publicly accessible

Balancing AI flexibility with deterministic safety was the hardest — and most rewarding — challenge.

What We Learned

  • AI is most effective when paired with deterministic systems
  • Security decisions must be explainable, not just accurate
  • Gemini 3 excels when used as a reasoning engine rather than a chatbot
  • Real-world AI systems require trust, transparency, and accountability

What's Next

We plan to extend SentinelAI with:

  • Browser extensions
  • Platform integrations (chat apps, email clients)
  • Multi-agent security workflows
  • Enterprise-grade analytics and dashboards

SentinelAI is a step toward secure, explainable AI-powered communication.

Built With

Share this project:

Updates