Uber

On this wiki, Uber appears as an enterprise operator and detection-engineering publisher rather than a security vendor. Its security and ML teams (with academic collaborators at MIT and the University of Oxford) built and deployed ADR, an agentic detection-and-response system for AI agents running over the Model Context Protocol, and released the ADR-Bench benchmark distilled from that deployment.¹

org_type is recorded as vendor for the closed-vocabulary lint, but the more accurate framing is enterprise practitioner: the value of Uber’s contribution is the production deployment evidence, not a product sold to others.

What the wiki cites it for

• Production deployment data. Ten months of ADR in production across 7,200+ unique hosts processing 10,000+ agent sessions daily on corporate macOS endpoints (Intel and ARM).¹ First session 2024-12-15; ≥100 sessions/day by ~April 2025; ≥10,000/day by ~October 2025.²
• Credential-exposure findings. Hundreds of high-severity credential exposures across 26 categories, and a Hooks-based shift-left prevention layer reported at 97.2% precision (206 true positives, 6 false positives across 212 unique credentials).²
• ADR-Bench. A 302-task, 133-MCP-server benchmark derived from Uber SOC telemetry, released open-source.¹

This makes Uber a peer production-scale data point to Salesforce’s Agentforce telemetry on the Agentic SOC axis — both report real volumes, real alert economics, and real false-positive rates rather than benchmark-only results.

See Also

• ADR — Agentic Detection for Enterprise AI
• ADR-Bench
• Agentic SOC: State of the Field

Footnotes

1. arXiv:2605.17380, authorship and §1, §4: Uber/MIT/Oxford authors, the ADR system, and ADR-Bench derived from enterprise telemetry. ↩ ↩² ↩³

2. §6 Real-World Deployment, arXiv:2605.17380: host and session counts, the deployment timeline (Figure 8), 26 credential categories, and the prevention-layer precision figures. ↩ ↩²