Predicting LLM Safety Before Release by Simulating Deployment

N/A

# link_1q9bmj2b2tl.pdf ## Metadata - PDFFormatVersion=1.5 - IsLinearized=false - IsAcroFormPresent=false - IsXFAPresent=false - IsCollectionPresent=false - IsSignaturesPresent=false - CreationDate=D:20260616174614Z - Creator=LaTeX with hyperref - ModDate=D:20260616112231-07'00' - Custom.PTEX.Fullbanner=This is pdfTeX, Version 3.141592653-2.6-1.40.27 (TeX Live 2025) kpathsea version 6.4.1 - Producer=pdfTeX-1.40.27 - Title=Predicting LLM Safety Before Release by Simulating Deployment - Trapped=False - dc:format=application/pdf - dc:title=Predicting LLM Safety Before Release by Simulating Deploym

saved by

• Justin Wang
• Cheikh Fiteni

related reading

• Predicting model behavior before release by simulating deployment | OpenAIopenai.com
• gpt-4.pdfcdn.openai.com
• Can public chat data predict real-world AI misalignments?alignment.openai.com
• Modifying LLM Beliefs with Synthetic Document Finetuningalignment.anthropic.com
• [2312.06942] AI Control: Improving Safety Despite Intentional Subversionarxiv.org
• 2312.06942arxiv.org
• AI Control: Improving Safety Despite Intentional Subversion — LessWronglesswrong.com
• gpt-4-system-card.pdfcdn.openai.com
• [2502.17424] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMsarxiv.org
• [2309.15817] Identifying the Risks of LM Agents with an LM-Emulated Sandboxarxiv.org
• [2510.13786] The Art of Scaling Reinforcement Learning Compute for LLMsarxiv.org
• [2506.17209] Fine-Tuning Lowers Safety and Disrupts Evaluation Consistencyarxiv.org