Anthropic’s Two New Power Moves: Desktop Takeover and Auto‑Approval Elimination

Anthropic rolled out two major updates within 48 hours: Claude Desktop’s Computer Use and Claude Code’s Auto Mode, each expanding the AI’s autonomy while adding layered safety controls.

Computer Use: From File Assistant to Full Desktop Control

Previously, Claude’s Cowork mode could only read and write files, create spreadsheets, and draft reports, but it could not interact with the broader desktop environment. With Computer Use, Claude can now capture the screen to perceive the current state and manipulate any application by controlling the mouse and keyboard. The implementation prefers API integration (e.g., Slack, Calendar) and falls back to screen‑based control only when no API is available.

A compelling scenario involves the Dispatch feature on a mobile device: a user texts “export that PPT to PDF and attach it to the 2 pm meeting invite,” and Claude silently performs all required clicks and keystrokes on the computer before the user reaches the meeting room.

Security is addressed through three layers: manual activation with user confirmation, per‑application authorization, and per‑application permission granularity (full control or view‑only). During operation, other windows are hidden to prevent accidental data exposure. The feature is currently limited to macOS for Pro and Max users and is labeled a Research Preview.

Auto Mode: Using AI to Review AI

Claude Code’s default mode requires the user to confirm every file write and shell command, which can become burdensome for large tasks. The previous workaround, the flag --dangerously-skip-permissions, explicitly warned of the risk.

Auto Mode introduces an independent classifier model (fixed to Sonnet 4.6) that performs a semantic safety review before each operation. The classifier evaluates the user’s intent and Claude’s planned action, determining whether the operation stays within the task scope or shows signs of malicious deviation.

Safe operations are automatically allowed, while dangerous ones are blocked; Claude then chooses an alternative way to continue its work. A notable design detail is that the results of tool execution are not sent back to the classifier, preventing prompt‑injection attacks that could influence safety decisions.

This feature is currently available to Team‑plan users, with Enterprise and API support slated to follow.

Shared Themes and Safety Philosophy

Both updates share a clear pattern: they broaden Claude’s autonomous action range while employing multi‑layer safeguards. Computer Use expands the operational space—from file‑only actions to full desktop manipulation—yet requires manual enablement, per‑app authorization, and granular permissions. Auto Mode expands the decision space—shifting from step‑by‑step user confirmation to self‑judgment—yet relies on an independent classifier that can downgrade to manual mode after repeated misjudgments.

The safety strategy is consistent: “don’t trust a single checkpoint; use layered checks.” Computer Use implements “enable confirmation → app authorization → permission granularity,” while Auto Mode follows “preset rules → automatic pass → classifier review.”

Anthropic’s broader narrative positions AI as an independent work partner rather than a passive tool. Computer Use gives the AI “hands,” Auto Mode provides it with “judgment,” and Dispatch enables remote “task assignment.” Combined, they outline a vision of an AI that can act autonomously on behalf of the user.