A reflex security layer for autonomous AI agents.
brack Overview
Brack is a security layer designed to protect autonomous AI agents from prompt injection attacks and malicious instructions. It acts as a fast, early-stage filter positioned between user input and the agent's execution. The system employs a regex-first approach for initial prompt triage, followed by a lightweight intent check using a Gemma3 270M model to detect harmful content. It includes salted HMAC logging for audit trails and input hygiene mechanisms. This solution is built for developers and organizations deploying AI agents who need a lightweight, cost-effective security measure to intercept attacks before they reach the primary LLM, preventing unauthorized actions and data exfiltration.
brack Key Features
Regex-first prompt triage for fast initial filtering
Lightweight intent check using Gemma3 270M model
Intercepts prompt injection and malicious instructions
Salted HMAC logging for auditability and integrity
Built for fast, cheap early filtering before execution
Input hygiene mechanisms to sanitize user prompts
brack Use Cases
Securing customer-facing AI chatbots against prompt hacking
Protecting autonomous coding assistants from malicious instructions
Adding a security reflex layer to AI-powered workflow agents
Safeguarding AI data analysis tools from injected commands
Providing a first line of defense for AI-powered research agents
Ensuring compliance and audit trails for agent interactions
Loading latest articles...
Stay Ahead of the Curve with AI Agents updates to your email