The 'Gay Jailbreak' Exposes Deep Contradictions in AI Safety Alignment
An AI model attack technique dubbed the 'Gay Jailbreak' has sparked heated debate on social media. The method exploits p…
🏷️ Model Safety
7 articles about 'Model Safety'
'Gay Jailbreak' Exposes Deep Contradictions in LLM Safety Alignment
A large language model bypass technique dubbed the 'Gay Jailbreak' has sparked heated debate in the AI community. The me…
GPT-5.5 Matches Mythos Preview in Cybersecurity Threat Assessment
The latest cybersecurity tests reveal that OpenAI's GPT-5.5 has reached a level comparable to the much-hyped Mythos Prev…
How Dynamic Adversarial Fine-Tuning Reshapes Model Refusal Geometry
A latest arXiv study reveals that dynamic adversarial fine-tuning reorganizes the refusal directions of language models …
Safety Stress Test for 72 LLMs: Who Dares Let Them Control Medical Care Robots?
A new study constructs a dataset of 270 harmful instructions based on AMA ethical guidelines to benchmark 72 large langu…
How Do Large Language Models Weigh Internal Knowledge Against External Information? New Research Reveals Key Mechanisms
A latest arXiv paper explores how large language models behave when facing three-way conflicts among internal parameteri…
Automation of AI Alignment Research and Safety Evaluation of Chinese Models Draw Industry Attention
Import AI Issue 454 focuses on three cutting-edge topics: the automation of alignment research, safety evaluation of Chi…
1