Sharing access control within a single set of model weights aims to make security compose across principals while utility may not, tested against held-out splits on a nested-factorisation mechanism. The authors…

#cybersecurity #AIafety #LLMsecurity #modelweights
https://arxiv.org/abs/2609.00445

LLM safety guardrails often reject prompts before generating any output, but researchers show harmfulness can be assessed from partial generation, reducing unnecessary refusals. This approach lets systems decide…

#AIafety #LLMSecurity #PromptEngineering #Guardrails
https://arxiv.org/abs/2610.16761

A new arXiv paper introduces GUISE, a cross-language benchmark showing that narrative wrapping boosts attack success against Qwen3-1.7B to 89.4% in English, 93.0% in modern Chinese, and 95.7% in Classical Chinese, with narrative…

#AIafety #LLMSecurity #RedTeaming
https://arxiv.org/abs/2610.11005

A voluntary White House pact with six major AI companies on safety safeguards raises more questions than it answers about real accountability. Without enforcement, it's…

#AIafety #Cybersecurity #AIGovernance #TechPolicy
https://www.wired.com/story/trumps-ai-safety-accord-is-a-fancy-pinky-swear/

SafeCoEvo proposes a test-time framework that co-evolves safety harnesses and guards for LLM agents, adapting from accumulated runtime experience rather than fixed task distributions.

#AIafety #LLMAgents #Cybersecurity
https://arxiv.org/abs/2609.36580

Geoffrey Hinton warns AI could pose a real threat to humanity, likening it to "alien beings." He calls for immediate global regulation as AI development accelerates.

Do you think AI's rapid advancement is concerning? #AIafety