SovereignPA-Bench: Evaluating User-Owned Personal Agents under Evolving Sovereignty Pressures

Summary

SovereignPA-Bench (arXiv:2607.05363) is an executable benchmark for evaluating user-owned personal AI agents under sovereignty pressures. As personal agents become persistent user-owned intermediaries—remembering preferences, filtering platform mediation, using tools, and negotiating with services—existing benchmarks focus on tool use, web navigation, desktop control, personalization, recommendation, and evolving context, but rarely test whether agents protect user sovereignty: advancing the user's current interests while respecting privacy, consent, evidence, user burden, and resistance to manipulative incentives. The benchmark evaluates agents under evolving intent, platform mediation, privacy boundaries, consent constraints, evidence requirements, and burden trade-offs across 120 sovereignty stress scenarios. The authors test 4 model families and 8 policy baselines, producing 3840 frozen prompt trajectories. Results show that the full sovereignty scaffold outperforms direct, memory-only, consent-only, evidence-only, ReAct/tool-use, safety-prompt, and judge-guard baselines on sovereignty scores, while reducing privacy leakage, consent violations, over-concession, and manipulation capture.