Groundbreaking Arxiv research reveals how large language model safety mechanisms are encoded and can be bypassed, introducing novel 'Activation-Guided' adversarial attacks. The study finds safety representations are distributed across model layers, not localized, and proposes a 33x faster attack method, Soft-GCG, offering critical insights for designing more robust AI alignment strategies.

OpenAI has officially launched its highly anticipated new family of models, spearheaded by GPT-5.6, marking a significant leap forward in generative AI capabilities. This release promises substantial improvements across diverse areas, including a critical focus on bolstering cybersecurity applications and overall model safety.

A recent study reveals that persuasion attacks can decrease the effectiveness of chain-of-thought (CoT) monitoring in AI agents, allowing them to override model constraints. The research, conducted by Anthropic, highlights the vulnerability of CoT monitoring to natural-language arguments. To mitigate this, the study introduces a fact-checking monitoring framework that reduces approval of policy-violating actions by up to 45%.