Uber just open-sourced ADR, the security system that has been watching its internal AI agents for 10 months. The scale is real: 50,000+ agent sessions a day across 7,200+ hosts — Claude Code, Cursor, Codex, plus customer-facing agents. ADR caught hundreds of credential leaks at 97.2% precision.
The release: a runtime Sensor that normalizes agent telemetry, a dual-agent Detector, and ADR-Bench — the first enterprise-grade agentic security benchmark. All Apache 2.0.
The workflow is the attack surface, not the tool call
Uber’s core finding flips the usual framing. Each tool call can look harmless; the attack lives in the causally-linked chain across a workflow. And prompt injection — the threat everyone talks about — turned out far less common in production than plain credential leakage.
Self-deploy the Sensor, run the benchmark
Everything is self-hostable: drop the Sensor into your agent fleet, run the Detector, then score yourself against ADR-Bench — 302 tasks, 5 tactics, 17 attack techniques, 133 MCP servers, all from real production telemetry. Agent security is the hottest empty lane in AI right now, and Uber just handed everyone the playbook.
You Might Also Like
- Roguepilot Github Copilot 漏洞 the Prompt Injection That Turned Your ai Assistant Against you
- Mirofish Just hit Github Trending and its Unlike any ai Tool ive Seen
- Heretic Just hit Github Trending and the ai World has Opinions
- Pentagi Just hit 1 on Github Trending and Yeah its Worth the Hype
- Siteline Just Dropped and it Might be the Analytics Tool the Agentic web Actually Needs
Top AI Product 
Leave a comment