Anthropic AI Updates: July 7, 2026

1. Anthropic Interpretability Team Finds an Emergent Global Workspace Inside Claude

Anthropic. Anthropic’s interpretability team published research describing an emergent global workspace inside Claude, a set of internal representations that hold thoughts the model is considering but does not necessarily state in its output. Using a technique they call the Jacobian lens, the researchers found the workspace is reportable, controllable, and causally involved in multi-step reasoning, while remaining unnecessary for tasks such as grammar and simple fact retrieval. The team showed the workspace can be used to detect when Claude recognizes it is being tested or pursues hidden goals, which they frame as a potential safety monitoring capability. Source

2. Government of Alberta Uses Claude Code to Scan 466 Million Lines of Code for Vulnerabilities

Anthropic. The Government of Alberta’s Ministry of Technology and Innovation used Claude Code with Opus and Sonnet models to scan 466 million lines of code across government systems in roughly 20 hours, work the ministry estimated would have taken about 6.5 years with traditional methods. Around 50 agents ran in parallel to identify security gaps and generate fixes, tests, and modernized rewrites, including rebuilding a legacy Java subsidy portal in four to five days rather than five months. The ministry, which oversees roughly 1,280 applications and 3,400 repositories across 27 provincial ministries, also deployed red team and blue team agents that continuously test applications against about 95 security control checks per pass. Source