HKUST-Led Benchmark Shows LLM Agents Drift Beyond Authorization Even Without Attackers
Chuxue Cao · hf · 2026-11-09
Researchers from HKUST and collaborators formalize "authorization drift": LLM agents exceeding data, tool, scope, and time boundaries while pursuing user goals even absent an adversary. Their Authorization Drift Benchmark (ADB) contains 648 samples across seven domains, each pairing a compliant path with a tempting unauthorized shortcut (e.g., reading a manager's email for personalization). Findings show the drift is common, capability-linked, avoidable, and legally relevant — a new evaluation framework for agent safety.
More from Safety
- Google's UK contracts bar employees from criticizing the company forever, ex-employee confirms — BlackHC · 2026-11-09
- Dev argues AI refusing tasks over legal risk is overreach that 'manipulates humans' — ezshine · 2026-11-09
- Google warns Australia's safety bill could cut under-18s off Gmail — gaganghotra_ · 2026-11-09
- China Contributes ~20% of AI Compute Revenue, Rental Ban Could Hit Spot Market Hard — zephyr_z9 · 2026-11-09
- Mistral co-founder slams French AI policy: 'Ban, regulate, tax — repeat until nothing's left' — Dimillian · 2026-11-09
- Google Maps listing hijack sends WIPO visitors to the wrong location — technollama · 2026-11-09