Skip to main content
AI News List

List of AI News about AI model safety

Time Details
2026-01-19
13:04 UTC
Anthropic Introduces Activation Capping to Counter Persona-Based Jailbreaks in AI Models

According to Anthropic (@AnthropicAI), persona-based jailbreaks exploit AI systems by prompting them to adopt harmful character roles, which can lead to unsafe responses. Anthropic has developed a new technique called 'activation capping' that constrains model activations along the 'Assistant Axis.' This method significantly reduces the likelihood of harmful outputs while maintaining the core capabilities and performance of the AI models. This advancement presents a practical solution for enterprises seeking robust AI safety mechanisms, especially for large language model deployment in regulated industries. Source: Anthropic (@AnthropicAI) on Twitter, Jan 19, 2026.

Source
2026-01-16
00:30 UTC
DeepMind’s Multi-Layered Constitutional Prompting: How Self-Correcting AI Principles Enhance Model Alignment

According to @godofprompt, DeepMind utilizes multi-layered constitutional prompting by applying a sequence of self-correcting principles to guide AI model outputs. Unlike public documentation that recommends being 'clear and specific,' DeepMind’s internal methodology requires models to verify compliance with one constitutional principle, revise if violated, and then iterate through additional principles. This rigorous process is designed to ensure that AI systems reason about and adhere to layered constraints, not just task completion, significantly improving model alignment, reliability, and safety in real-world AI applications (source: @godofprompt on Twitter, Jan 16, 2026).

Source
2025-12-09
11:47 UTC
SGTM: Selective Gradient Masking Enables Safer AI by Splitting Model Weights for High-Risk Deployments

According to Anthropic (@AnthropicAI), the Selective Gradient Masking (SGTM) technique divides a model’s weights into 'retain' and 'forget' subsets during pretraining, intentionally guiding sensitive or high-risk knowledge into the 'forget' subset. Before deployment in high-risk environments, this subset can be removed, reducing the risk of unintended outputs or misuse. This approach provides a practical solution for organizations seeking to deploy advanced AI models with granular control over sensitive knowledge, addressing compliance and safety requirements in regulated industries. Source: alignment.anthropic.com/2025/selective-gradient-masking/

Source
2025-12-09
11:47 UTC
SGTM AI Unlearning Method Proves More Difficult to Reverse Than RMU, Reports Anthropic

According to Anthropic (@AnthropicAI), the SGTM (Stochastic Gradient Targeted Masking) unlearning method is significantly more resilient than previous approaches. Specifically, it requires seven times more fine-tuning steps to recover forgotten knowledge using SGTM compared to the RMU (Random Masking Unlearning) method. This finding highlights a critical advancement for AI model safety and confidential data retention, as SGTM makes it much harder to reintroduce sensitive or unwanted knowledge once it has been unlearned. For enterprises and developers, this strengthens compliance and data privacy opportunities, making SGTM a promising tool for robust AI regulation and long-term security (source: Anthropic, Twitter, Dec 9, 2025).

Source