Shortcut Sensitivity in Agent-Safety Evaluations
AI agents take actions through tools.
ProofPaper ↗
Key points
- We introduce a three-level shortcut-auditing framework and audit completed LinuxArena trajectories and two TS-Bench-derived step datasets.
- We distinguish label predictability, observational guard association, and paired decision sensitivity.
- In AgentHarm-Traj, released TS-Guard F1 falls from .902 under row weighting to .719 with equal interaction weighting, while balanced accuracy rises.
- We then test the strongest safe-row marker association in a seven-configuration screen on a frozen 64-step AgentDojo-Traj cohort using repeated, interleaved Haiku calls.
Sources (1)
- [1]Shortcut Sensitivity in Agent-Safety EvaluationsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 05:18 PM
AI agents take actions through tools.
We introduce a three-level shortcut-auditing framework and audit completed LinuxArena trajectories and two TS-Bench-derived step datasets.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026Friends don’t let friends feed their data to AI agents
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026Code, Test, and Deploy on One Machine with NVIDIA RTX Spark
- Oct 15, 2026Matoi: Selective Revalidation of Consequence Evidence for Agent Authorization
- Oct 15, 2026Approve What, Exactly? Effect Binding and Trusted Rendering for Agent Step-Up Authorization
- Oct 14, 2026Floqer’s AI agents research sales prospects so teams don’t have to