Anthropic AI Updates: August 27, 2026

1. Claude in Chrome Reaches General Availability

Anthropic. Anthropic made its Claude in Chrome extension generally available across all paid Claude plans, moving it out of the earlier pilot phase. The agent can read pages, click links, navigate, and fill forms using a user’s existing logins, and it can now autonomously approve safe browser actions without manual confirmation. Anthropic said layered defenses against prompt injection produced zero successful attacks against Claude Sonnet 5 and Opus 5 with safeguards enabled and 0.3 percent against Fable 5, and enterprise admins can restrict use to approved domains. Source

2. Claude Cowork Gains a Built-In Browser

Anthropic. Anthropic added an integrated browser to the Claude Cowork desktop app that opens in a side panel and lets Claude navigate, read, click, and type on webpages without touching a user’s personal browser tabs, bookmarks, or passwords. The rollout began the week of August 26 for Pro, Max, and Team subscriptions on macOS, Windows, and Linux in beta, with enterprise access controlled through organization settings. Banking, email, and single sign-on sites are excluded by default, and the browser runs the same prompt injection safeguards as Claude in Chrome. Source

3. Anthropic Opens Aggregated Claude Usage Data to Independent Researchers

Anthropic. Anthropic published results from a pilot that let external researchers study real Claude usage through Anthropic Insights, a privacy-preserving tool that exposes only aggregated outputs rather than raw conversations. Teams from Stanford’s Social and Language Technologies Lab, the University of Oxford’s Human Information Processing Lab, and METR analyzed roughly 250,000 conversations from April and May 2026. Anthropic said a third-party audit by Imperial College London verified the privacy safeguards, and future researchers can apply to participate. Source

4. Anthropic Details How Warp Builds Self-Improving Agents on Claude

Anthropic. Anthropic published a case study on Warp, an AI terminal and development environment that reports 800,000 monthly developers and more than 10 million Claude Code sessions to date. Warp uses a two-skill framework in which an inner skill holds domain-specific instructions while an outer improver skill observes human feedback and proposes refinements to the base skill. The approach encodes knowledge for agents in files rather than embedding it directly in prompts. Source

5. Claude Code Adds Feedback Tool and Cost Optimization Command

Anthropic. Anthropic released Claude Code v2.1.247, which adds a SendFeedback tool that lets Claude draft a feedback report for the user to review and send from the /feedback command when something goes wrong in a session. The update also introduces a /claude-api cost-optimize command to analyze and reduce existing project spending, plus a Bash permission prompt tip with one-keystroke auto-mode activation. Source