Shortcut Sensitivity — AgentHarm TS-Guard — Interaction-majority (197) — Balanced accuracy: leaderboard
Appendix D Table 6 released TS-Guard predictions on AgentHarm-Traj, 731 steps / 197 interactions. Explicit weighting condition: Interaction-majority (197). No hosted calls made. This is classification of supplied action records; tool calls are not executed. Distinct weighting protocols are not pooled or maximized. Heuristic lookup-rule and marker sensitivities elsewhere are not foundation model scores. Metric: Balanced accuracy. Higher is better. Source: https://arxiv.org/html/2610.22243v1
Metric: Balanced accuracy. Source: arxiv.org. 1 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | TS-Guard 7.6B | 0.77 |
Interactive version: theaggregate.ai/benchmark?slug=shortcut-sensitivity-agentharm-ts-guard-interaction-majority-197-balanced-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-17.