Uber just open-sourced their AI agent security stack. This is what every company running agents in prod needs right now.
They call it ADR — basically a security layer watching every AI agent employees touch ($CLAUDE, Cursor, MCP servers, etc.)
What it does:
• Discovery — auto-detects every agent/app deployed
• Observability — logs agent actions + reasoning
• Detection — catches sus agent behavior in real-time
• Benchmark — 300+ test cases covering 17 attack vectors
This isn't a whitepaper flex. It's running live at Uber right now. Accepted to MLSys 2026. Apache 2.0 license.
If you're shipping agents at scale, you need this layer. Agents without guardrails = liability.
Repo is live.
They call it ADR — basically a security layer watching every AI agent employees touch ($CLAUDE, Cursor, MCP servers, etc.)
What it does:
• Discovery — auto-detects every agent/app deployed
• Observability — logs agent actions + reasoning
• Detection — catches sus agent behavior in real-time
• Benchmark — 300+ test cases covering 17 attack vectors
This isn't a whitepaper flex. It's running live at Uber right now. Accepted to MLSys 2026. Apache 2.0 license.
If you're shipping agents at scale, you need this layer. Agents without guardrails = liability.
Repo is live.