Modular Knowledge Compartmentalization for AI Models

4 posts · Aug 2025 – Jul 2026

What this trail covers

Anthropic is developing and iterating on novel methods for technical candidate evaluation that remain effective and distinguishing in the face of advancing AI agent capabilities. This involves a continuous process of designing, testing, and iterating on evaluation formats, such as automated evals and complex cyber evaluations, to ensure they accurately assess agent capabilities and behaviors. Recent iterations have involved increasingly complex problem spaces, multi-turn evaluations, and the dev

All Anthropic trails

Jul 2026An off switch for dual use knowledge in AI modelsIntroduces Gradient-Routed Auxiliary Modules (GRAM), a novel method for training AI models with modular, removable compartments for dual-use knowledge. GRAM adds auxiliary neurons grouped into modules per dual-use category. During training, only the relevant module learns from dual-use data, while general weights are frozen. This allows for selective deletion or retention of capabilities without retraining separate models. Experiments show GRAM effectively isolates dual-use knowledge (virology, cybersecurity, nuclear physics, specialized programming language) and removes it without degrading general performance, matching data filtering effectiveness across various model sizes (50M to 5B parameters). GRAM also demonstrated resistance to attacker attempts to recover removed knowledge via fine-tuning.
Jan 2026Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaksIntroduced Constitutional Classifiers++ (CC++), an evolution of the Constitutional Classifiers system for defending against LLM jailbreaks. CC++ features a two-stage cascade architecture: a lightweight probe classifier screens internal model activations for suspicious patterns, and if flagged, escalates to a more powerful probe-classifier ensemble that monitors both inputs and outputs. This system achieves a 0.05% refusal rate on harmless queries and a detection rate of 0.005 per thousand queries, with no universal jailbreaks discovered during extensive red-teaming. The system adds approximately 1% compute overhead.
Sep 2025LLMs and bioriskThis post details Anthropic's ongoing efforts to address biorisk associated with AI models. It expands on the ASL-3 protections activated for Claude Opus 4, specifically concerning biological weapons development. The post highlights the striking prioritization of biorisk across frontier AI labs, despite LLMs being generalists. It discusses the dual-use nature of AI, comparing its potential misuse to historical information technologies while emphasizing AI's unique capability to act as a true assistant. The post presents evidence of LLMs nearing or exceeding expert human performance in biological knowledge and outlines controlled trials demonstrating AI's superior assistance in planning hypothetical bioweapons acquisition. It also details preliminary and ongoing wet lab studies, co-sponsored with the Frontier Model Forum and Sentinel Bio, to investigate AI uplift in real laboratory tasks and the role of tacit knowledge.
Aug 2025Developing Nuclear Safeguards for AICo-developed an AI classifier with the NNSA and DOE national laboratories to distinguish concerning nuclear-related conversations from benign ones with 96% accuracy in preliminary testing. This classifier has been deployed on Claude traffic to monitor for misuse. The approach involved using a curated set of nuclear risk indicators shared by the NNSA to build a real-time content classifier. Validation was performed using synthetic data generated from NNSA examples, and the system was refined based on feedback. Real-world deployment revealed the need for hybrid safety systems, such as hierarchical summarization, to handle nuanced conversations and reduce false positives. This work demonstrates a model for public-private partnerships in AI safety and risk mitigation.
Advertisement