Anthropic Develops Novel System to Restrict Dual-Use AI Risks
The post Anthropic Develops Novel System to Restrict Dual-Use AI Risks appeared on BitcoinEthereumNews.com. Timothy Morano Jul 08, 2026 23:24 Anthropic introduces GRAM, a method to control dual-use AI knowledge, amid policy concerns over unchecked AI risks. Anthropic has unveiled GRAM (Gradient-Routed Auxiliary Modules), a novel approach to controlling access to dual-use knowledge in AI models, according to a research update published on July 8, 2026. Dual-use knowledge refers to AI capabilities that can serve beneficial purposes, such as cybersecurity or virology research, but can also be weaponized for malicious intents. GRAM aims to surgically limit access to such knowledge without requiring separate, costly retraining for different use cases. Current dual-use safeguards, like refusal training and classifiers, often fail to provide robust protection. These methods block potentially harmful outputs but do not address the knowledge embedded in the model itself. GRAM,