Uber's Persona Guardrail lifts agent defense accuracy to 95.9%, cuts false approvals to 4.7%

amaarora · x · 2026-10-07

Uber Engineering published Persona Guardrail, a production-deployed runtime defense framework for agentic AI systems, along with the PAGE benchmark.

Core idea: Existing guardrails mostly target specific attacks like prompt injection and can't ensure agents stay within intended functionality. Persona Guardrail enforces explicit functional boundaries via semantic allowlist/blocklist specifications, with synchronous input and output validation for customer-facing agents.
PAGE benchmark: Evaluates function-specific guardrails across benign, adversarial, and out-of-domain interactions on both user and agent turns.
Results: Versus a generic LLM-based guardrail, overall accuracy improves from 85.7% to 95.9%, out-of-domain detection jumps from 57.3% to 93.5%, and the false-approved rate drops from 25.0% to 4.7% — all within its latency budget.

More from coding & agent