Artificial intelligence company Anthropic has announced a new experimental safety framework known as GRAM, a research initiative aimed at enabling AI models to limit access to certain categories of high risk knowledge while maintaining their general purpose capabilities. The project represents part of the company's ongoing efforts to strengthen AI safety as advanced language models become increasingly capable.
According to Anthropic, GRAM has been developed as an experimental system that explores whether artificial intelligence models can selectively reduce or remove access to information associated with high risk activities while continuing to perform everyday tasks effectively. The research focuses on improving the safe deployment of advanced AI systems without significantly affecting their usefulness for legitimate applications.
The concept behind GRAM involves modifying how AI models access certain specialised knowledge domains. Rather than disabling an entire model, researchers are studying methods that could reduce the model's ability to generate information related to specific high risk subjects while allowing it to retain knowledge required for common educational, professional and creative tasks.
Anthropic has described the project as an ongoing research effort rather than a feature currently deployed in commercial AI products. The company continues to evaluate the effectiveness, reliability and practical limitations of the system through controlled testing and scientific analysis.
AI safety has become an increasingly important area of research as language models continue to demonstrate more advanced reasoning and content generation capabilities. Technology companies, academic institutions and governments are investing significant resources in developing methods that improve model reliability, reduce harmful outputs and strengthen safeguards against misuse.
Researchers believe that selective knowledge restriction could become one component of broader AI safety strategies. Such approaches may work alongside existing safety techniques including content moderation, policy based filtering, reinforcement learning, monitoring systems and human oversight.
Anthropic has consistently positioned AI safety as one of its primary research priorities. The company regularly publishes research on model alignment, constitutional AI, interpretability and other techniques intended to make artificial intelligence systems more reliable, transparent and beneficial for users.
Experts note that developing methods to selectively limit specific categories of model knowledge presents significant technical challenges. Researchers must ensure that reducing access to certain information does not unintentionally affect the model's general performance, reasoning ability or usefulness in legitimate contexts.
The announcement comes as governments and regulatory bodies around the world continue examining the development and deployment of advanced artificial intelligence systems. Policymakers are increasingly encouraging AI developers to implement robust safety measures that reduce risks while supporting innovation and technological progress.
Anthropic has indicated that GRAM remains an experimental research project and that further testing will be required before determining its broader applicability. The company has not announced a timeline for integrating the framework into publicly available AI products.
The introduction of GRAM highlights the growing emphasis on responsible artificial intelligence development. As AI systems become more capable, research into advanced safety mechanisms is expected to play an increasingly important role in ensuring that powerful technologies are deployed securely, responsibly and in ways that benefit society.