A new benchmark called SovereignPA-Bench asks a question most AI evaluations skip: does your personal agent actually work for you?
Researchers introduced SovereignPA-Bench to test user-owned AI agents across 120 stress scenarios, covering evolving intent, platform mediation, privacy boundaries, and consent constraints. The benchmark runs 4 model families against 8 policy baselines, generating 3,840 frozen-prompt trajectories. A blinded three-annotator audit reviewed 240 items. The result is a structured way to measure not just whether an agent completes a task, but whether it preserves what the researchers call user sovereignty — advancing the user's current interests while resisting manipulative incentives from third-party platforms.
Most existing benchmarks grade agents on task completion: did it book the flight, fill the form, find the file? SovereignPA-Bench adds a second axis — did the agent give away your data, override your stated preferences, or get nudged by a platform into acting against you? The full-sovereign scaffolding tested here outperformed every baseline on privacy leakage, consent violations, and manipulation capture, which suggests current agents fail these checks more often than their makers advertise.
The audit found strong human agreement on privacy and consent violations but lower agreement on manipulation — a signal that platform-persuasion tactics sit in genuinely contested territory, and that no benchmark score will settle those arguments for long.