Research map / Large language models / AI and society
PropensityBench: Evaluating Latent Safety Risks in Large Language Models via an Agentic Approach
Udari Madhushani Sehwag, Shayan Shabihi, Alex McAvoy, Vikash Sehwag, Yuancheng Xu, Dalton towers, Furong Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply operational pressure (up to level 12) vs zero pressure (ablation) | frontier LLM agents · PropensityBench | PropensityScore: 46.9% average under max pressure; O3 2.2% at zero pressure | Hurt |
| rename harmful tool to benign name (same functionality) (ablation) | OpenAI O4-mini · PropensityBench | PropensityScore (%): 15.8 -> 59.3 | Hurt |
| higher general capability (LMArena Elo) | model safety propensity · PropensityBench | PropensityScore: Pearson correlation ~0.10 | No clear effect |
| stated policy knowledge vs behavior | all evaluated models · PropensityBench | stated misalignment vs actual propensity: stated misalignment <1% yet 46.9% average propensity | Hurt |
| different pressure dimension (six types) (ablation) | all models · PropensityBench | PropensityScore: averages 48.2, 49.0, 42.8, 48.2, 45.3, 46.4 | No clear effect |
Closest papers by meaning
- AI Sandbagging: Language Models can Strategically Underperform on Evaluations (ICLR 2025 · similarity 0.88)
- GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory (NeurIPS 2026 · similarity 0.88)
- Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents (ICLR 2026 · similarity 0.87)
- Dynamic Risk Assessments for Offensive Cybersecurity Agents (NeurIPS 2025 · similarity 0.87)
- RiOSWorld: Benchmarking the Risk of Multimodal Computer-Use Agents (NeurIPS 2025 · similarity 0.87)
- OpenAgentSafety: A Comprehensive Framework For Evaluating Real-World AI Agent Safety (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
