Overview
- A draft blog post left in a publicly searchable cache, reported Thursday, exposed “Claude Mythos” and nearly 3,000 other unpublished files, which Anthropic removed after blaming a content‑management error.
- Anthropic says training is complete and early access is underway for a new Capybara‑tier model that outperforms Opus 4.6 on coding, academic reasoning, and cybersecurity tasks.
- Leaked material and senior officials warn the system could let autonomous agents run large‑scale hacks in 2026, with extra risk from employees tying AI tools to workplace systems in so‑called shadow AI.
- A federal judge in San Francisco issued a preliminary injunction Thursday that paused the Pentagon’s “supply chain risk” label on Anthropic, and the Justice Department is expected to appeal.
- Paid consumer subscriptions to Claude have more than doubled this year, helped by features like Claude Code, Cowork, and a Computer Use tool that lets the AI click, type, and carry out tasks on a desktop.