BlackGrid

Inspiration

Modern security systems are built around static rules and known attack patterns, but threats are no longer predictable. As autonomous agents become more capable, they can chain tools, mimic real users, and hide malicious intent inside normal workflows.

Traditional tools can surface logs, but they cannot understand intent. This creates a critical gap where semantic attacks like persona spoofing, context manipulation, and protocol abuse go undetected.

We wanted to build a system that goes beyond scanning—one that understands behavior, interprets intent, and responds in real time.

A simple idea inspired BLACKGRID: what if you could detect threats the way a human analyst would—by understanding what an agent is trying to do, not just what it outputs?

What it does

BLACKGRID is a cybersecurity platform that detects agentic threats, analyzes semantic telemetry, and autonomously contains attacks in real time.

Captures raw telemetry and execution traces from agent workflows
Interprets intent using semantic embeddings instead of keyword matching
Scores threats based on behavioral similarity to known attack patterns
Simulates attacks using realistic, context-aware telemetry generation
Executes containment through deterministic response actions

In seconds, users go from telemetry → intent → containment.

How we built it

We built BLACKGRID as a hybrid system that separates threat simulation from detection, enabling realistic testing while keeping defense fast and local.

Defense Engine (Local / Edge):

TensorFlow.js running fully in the browser (WebGL accelerated)
Universal Sentence Encoder converts telemetry into high-dimensional embeddings
Vector similarity used to detect semantic alignment with known threat behaviors
Enables detection of novel threats without predefined signatures

Attack Simulation (Cloud):

Gemini API generates high-fidelity, context-aware attack logs
Simulates behaviors like tool chaining, context overflow, and protocol abuse
Supports both dynamic (cloud) and offline (local template) generation

Frontend & Visualization:

React + TypeScript for a structured command interface
Recharts for real-time visualization (threat scores, system activity)
Tailwind CSS for a clean, system-level UI

This architecture enables real-time, privacy-preserving analysis with high-fidelity threat simulation.

GitHub: https://github.com/AbhinavGGarg/BlackGrid

Challenges we ran into

Running neural models efficiently in the browser while maintaining low latency
Managing model load times and optimizing TensorFlow.js performance
Generating realistic attack telemetry that reflects modern agent behavior
Designing a system that balances technical depth with usability

Accomplishments that we're proud of

Built a fully browser-based system for real-time semantic threat detection
Achieved fast, low-latency analysis using local neural inference
Designed a hybrid architecture combining simulation and on-device defense
Enabled detection of novel threats through intent-based analysis
Delivered an end-to-end system from telemetry ingestion to containment

What we learned

Local inference can be both fast and privacy-presing when optimized correctly
Semantic analysis is more effective than rule-based detection for modern threats
Agent-based systems introduce new attack surfaces requiring behavior-first security
Strong UX is critical for making complex systems understandable and actionable