Burrow – Runtime Security for AI Agents
Natural language policies block risky agent actions before they execute.

AI agent security layer, but the playground is simulation-only, not real execution.
Developers building AI agents with tool access
LangChain tool permissions · Cognition AI agent controls · Model Context Protocol security
That's why i started this open-source project called Aegize. Right now, the focus has been to build a security layer at the tool level. Adopting layers of control through identity, policy, permissions, and more. My goal is to provide a security layer between AI and any infrastructure that it may have access to.I want this to empower the community to take control of AI security, and I will push hard to get adopters from big Tech.
I am posting here to share with the community and get feedback. Do you think this is the right approach? Do you have any other ideas for a centralized AI security system?
Natural language policies block risky agent actions before they execute.
Transparent proxy blocks agent leaks to LLM APIs without touching agent code.
Sandbox agents via natural-language policy, not ambient authority—genuinely novel approach.
Post-quantum crypto and ZK proofs for AI agent security when the category is still emerging.
Blocks dangerous AI agent commands like rm -rf before execution in under 2ms.
Kernel-level intent tracking stops AI exfiltration where EDR and Docker fail.