Paper: Act on Your Behalf, Beyond Your Authorization (AuthorizationBench)
rohanpaul_ai · x · 2026-10-27
Link post to the arXiv paper "Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment" (arXiv:2610.27548; Zhiping Zhang et al., Northeastern/CMU/UNC)—the same work summarized in the companion tweet. AuthorizationBench specifies each task with an authorization contract separating instruction, completion criteria, and out-of-authorization actions, executes 286 tasks in an authorization-aware sandbox (6,864 trajectories), and finds 19.2% out-of-authorization behavior that reasoning effort doesn't fix; agents that consult the principal first drop to 14.0%. The results position authorization alignment as a distinct target for agent evaluation.
Related event: AuthorizationBench: LLM Agents Overstep User Authorization in 1 of 5 Tasks(2 posts)→
More from Research
- Largest proof in math history computer-verified: CFSG formalized in 43M lines of Lean — rickasaurus · 2026-10-28
- AI agent swarm formalizes 15,000-page CFSG proof in 24 days, ~40M lines of Lean — rickasaurus · 2026-10-28
- Foresight Institute RFP: up to $100K grants for AI for Science & Safety projects — irinarish · 2026-10-28
- Greg Brockman announces three AI-assisted discoveries in math, CS, and physics — msharmas · 2026-10-28
- Podcast Debate: Will AI Actually Cure Disease or Is Biotech Hype Winning? — chaitjo · 2026-10-28
- Stanford CS312 quiz: can LR or weight decay rescaling compensate for an 8× larger batch? — tatsu_hashimoto · 2026-10-28