Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety

arXiv:2606.06529Submitted Jun 8, 20260 benchmark results

Authors pending

Abstract

Strategic start/stop attack policies reduce measured safety by 20–28pp at 1% audit budget on BashArena and LinuxArena, suggesting control evaluations may be overly optimistic.

Tasks

edit

• Agentic AI

Results

No benchmark results recorded yet.

submit

Benchmark results referencing this paper haven't been added to the registry yet. If you have a reproduction, submit it →

CodeSOTA extraction

Benchmark evidence

edit

Attack Selection: exact safety reduction in pp at 1% audit (abstract reports 20pp on BashArena and LinuxArena start policy)

Add or update benchmark results

Logged-in editor · benchmark trail

Three places to go from here.

Index

All papers

All tracked papers in the registry, with benchmark result, model, and leaderboard linkage where available.

Replacement

Papers with Code is dead — alternatives

What replaced PWC for each use case: LLMs, OCR, speech, vision, robotics.

Top hub

Agentic AI

Every benchmark in Agentic AI.