Anthropic AI Updates: July 9, 2026

1. Anthropic Introduces GRAM, an Off Switch for Dual-Use Knowledge in AI Models

Anthropic. Researchers unveiled GRAM (Gradient-Routed Auxiliary Modules), a technique that adds dedicated, removable compartments to a Transformer so dual-use knowledge in areas like virology, cybersecurity, and nuclear physics can be isolated and later deleted. In testing, removing a module eliminated the corresponding capability about as effectively as never training on that data, without degrading general performance. The team stresses the work is preliminary and has not been applied to production Claude models or tested at frontier scale. Source

2. Thomson Reuters Builds “Fiduciary-Grade AI” for High-Stakes Legal Work With Claude

Anthropic. Anthropic detailed how Thomson Reuters integrates Claude into CoCounsel Legal and other products, combining its models with authoritative content and workflow tooling in an approach it calls “Fiduciary-Grade AI.” The company rebuilt CoCounsel Legal on the Claude Agent SDK to orchestrate hundreds of tools in real time and prioritized citation verification so professionals can review the output. CTO Joel Hron argues verification matters more than fluency because the human professional remains accountable for the final work product. Source

3. Anthropic’s Marketing Team Automates Reporting With Claude Cowork

Anthropic. Anthropic described how its own marketing operations team uses Claude Cowork to automate repetitive reporting and campaign-setup workflows. One team member cut a weekly metrics review from one to two days down to roughly two hours, while another automated event infrastructure setup across CRM, marketing automation, and event tools using a dispatcher skill that routes requests to specialized skills. The shift freed both to focus on validation rather than manual data gathering. Source