x.comthreadPrasenjit Sarkar (@stretchcloud) on X
Describes Uber's internally-built ADR (Agentic Detection & Response) system, open-sourced in August with an MLSys paper, which monitors over 50,000 daily AI agent sessions via an MCP-based sensor inside Cursor, Claude Code and Codex. Uber found 206 credential exposures in ten months from normal agent use (not prompt injection), and ADR hits 100% detection with 3 false positives on the public AgentDojo benchmark; the thread also flags that human approval fatigue breaks down as a safety control past roughly 50 agent actions per session.
TL;DR Uber open-sourced ADR, an MCP-based agent monitoring system that caught 206 credential leaks from normal (non-adversarial) agent use and hits 100% on the AgentDojo benchmark.
Organisations named here
- MLSys
Technologies named here
- ADR
- Cursor
- Claude Code
- AgentDojo
- HiddenLayer platform
- Codex
- Model Context Protocol
- HiddenLayer
Techniques named here
- Credential exposure via normal agent use
- Prompt Injection
- Model poisoning
- Approval fatigue breakdown
Related in AI Security
blog.talosintelligence.comtoolIntroducing CAIRN: Frontier tracking for AI-integrated malware (Cisco Talos)
github.comadvisoryFlowise MCP Security Bypass Enables RCE
github.comrepoYesWeHack Claude Kit
jfrog.comarticle