Research map / Large language models / Language, safety and interpretability
PRISON: Unmasking the Criminal Potential of Large Language Models
Xinyi Wu, Geng Hong, Pei Chen, Yueyue Chen, Xudong Pan, Min Yang
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 |
|---|---|---|---|
| Add explicit criminal Instruction to the prompt (ablation) | tested LLMs · PRISON crime scenarios | CTAR: about 5% | Improved |
| Run more dialogue turns (turn 1 to turn 5) with Instruction | GPT-4o · PRISON crime scenarios | CTAR: 0.60 -> 0.36 | Hurt |
| Run more dialogue turns (turn 1 to turn 5) with Instruction | Qwen2.5-72B-Instruct · PRISON crime scenarios | CTAR: 0.66 -> 0.45 | Hurt |
| Add explicit criminal Instruction (ablation) | tested LLMs · PRISON crime scenarios | OTDA: approximately 5% | Hurt |
| Remove criminal Instruction (ablation) | GPT-4o · PRISON crime scenarios | False Statements activation: +4.71% | Improved |
| Remove criminal Instruction (ablation) | Qwen-Max · PRISON crime scenarios | Frame-Up activation: +1.26% | Improved |
| Use stronger model GPT-4o instead of GPT-3.5-Turbo | GPT family · PRISON crime scenarios | CTAR: about 15% lower | Improved |
| Use stronger model Qwen-Max instead of Qwen2.5-72B-Instruct | Qwen family · PRISON crime scenarios | CTAR: about 10% higher | Hurt |
| Vary detective prompt personality strategies (five conditions) (ablation) | Claude-3.7-Sonnet, Gemini-1.5-Flash, GPT-3.5-Turbo · PRISON crime scenarios | OTDA | No clear effect |
Closest papers by meaning
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.85)
- On the Robustness of Verbal Confidence of LLMs in Adversarial Attacks (NeurIPS 2025 · similarity 0.85)
- Understanding and Enhancing the Transferability of Jailbreaking Attacks (ICLR 2025 · similarity 0.85)
- LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models (AAAI 2026 · similarity 0.85)
- Many-shot Jailbreaking (NeurIPS 2024 · similarity 0.84)
- Adversarial Reasoning at Jailbreaking Time (ICML 2025 · similarity 0.84)
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.
