Anthropic and AE Studio have introduced GRAM (Gradient-Routed Auxiliary Modules), a new technique for giving AI models dedicated, removable compartments for dual-use knowledge such as cybersecurity or virology, allowing developers to toggle specific knowledge domains on or off by adding and removing targeted modules. Early experiments show that removing a module effectively erases the associated capability with little impact on general performance, but GRAM has not yet been tested at frontier scale or in production pipelines, and researchers caution that some dual-use knowledge may be too entangled to isolate cleanly. According to Futurum Group's 1H 2026 AI Platforms Decision Maker Survey (n=820), 53% of organizations cite privacy and security as top GenAI adoption challenges, just behind reliability and hallucination management at 55%.