What AI Update Did You Miss Today?

Claude Code Adds an Auto Mode Tab to /permissions

Anthropic added a dedicated Auto mode tab to Claude Code's /permissions command, letting users view and edit the classifier rules that govern which actions auto mode approves without prompting. The tab surfaces the allow, soft-deny, hard-deny, and environment rules pulled from every settings scope Claude Code reads, and shows whether the built-in safety rules remain active for each section. Rules added through managed settings or the --settings flag appear as read-only, while everything else can be edited directly. The change ships in Claude Code 2.1.246 and gives auto mode users a visual alternative to hand-editing JSON to tune what the classifier lets through.

Claude Code

Claude Code Patches Six Permission, Sandbox, and Isolation Gaps

Claude Code 2.1.246 closes six separate gaps in how the tool enforces permissions, sandboxing, and session isolation. Fixes include Bash commands with a dangling && or || operator that previously skipped approval, sessions that lost plan mode's execution restrictions after resuming, a sandbox filesystem check that ignored the --setting-sources flag, and cross-session leaks where one launch's uncommitted changes appeared in another. Anthropic also patched a credential leak that sent a third-party gateway's API key to its own telemetry endpoint, and added a startup warning for Bash allow rules whose wildcard sits before the subcommand, since those rules can match more than intended.

Claude Code

Claude Code: Sign In With Your Anthropic Console Account

Claude Code added a keyless sign-in option under `/login` that lets people authenticate with their Anthropic Console account instead of creating an API key. Anthropic built the option for organizations whose policies don't allow individual API keys, giving them a browser-based OAuth path into Claude Code. The addition sits alongside the existing "create an API key" flow rather than replacing it, and appears in version 2.1.243, released August 24, 2026.

Claude Code

Claude Code Makes Sonnet 5's Introductory Pricing Permanent

Claude Code's `/model` picker and its bundled `claude-api` skill now show Claude Sonnet 5's $2-per-million-input / $10-per-million-output token pricing as the standard list price rather than a limited-time promotion. Anthropic had originally introduced that rate in June 2026 as a promo running through August 31, with a published increase to $3/$15 planned for September 1; the company canceled that increase and made the lower rate permanent. The update in Claude Code 2.1.243 brings the CLI's own pricing displays in line with that decision.

Claude Code

Claude Code Cuts Install Size and Memory Use for Native Builds

Anthropic significantly shrank Claude Code's native install and auto-update footprint, compressing the binary with zstd to cut the Linux x64 download from roughly 340 MB to about 75 MB. The same release trims the bundled binary further by storing skill and prompt text more compactly, and lowers runtime memory use by roughly 40 to 70 MB per session by loading code on demand instead of keeping the whole bundle resident. Startup was also made faster by moving sandbox and MCP bring-up off the critical path for the first frame.

Claude Code

Claude Code Closes a Computer-Use Permission Gap on macOS

Claude Code's macOS computer-use feature previously let Claude click the desktop, the Dock, or a Finder window without first prompting for the same per-app access grant required for every other application. Claude Code 2.1.243 closes that gap: interacting with the desktop, Dock, or Finder now requires the standard Finder access-dialog grant, matching how every other app is handled. This is a trust-boundary fix, since it removes a path where Claude's on-screen actions were not gated by the permission prompt users rely on to control what it can touch.

Claude Code

Claude Code Patches a Web Sandbox Proxy Bypass

Anthropic fixed a network sandbox gap in Claude Code on the web, where requests from Bash and other tools to non-API anthropic.com hosts such as www and docs bypassed the session's network proxy. Because environment allowed-domain restrictions are enforced at that proxy, traffic to those hosts previously escaped the allowlist check meant to govern every outbound connection from a cloud session. Version 2.1.239, released August 21, 2026, routes these requests through the proxy like any other outbound call, so the environment's configured allowed domains now apply consistently. The fix is the latest in a recurring pattern of narrow gaps found in Claude Code's network sandbox enforcement.

Claude Code

Claude Code Patches Four Auto Mode and Bash Permission Gaps

Across versions 2.1.236 and 2.1.238, Anthropic closed four separate gaps in Claude Code's Auto Mode and Bash tool permission checks: Monitor commands skipping classifier review, a weaker classifier on Bedrock/Vertex/Foundry deployments, a `git status`-spoofing trick, and a zsh-specific syntax gap in the Bash permission checker. None is a dramatic standalone exploit, but together they follow the same "small cracks in the approval layer" pattern as Anthropic's earlier batch of ten permission and sandbox fixes. The timing matters: Auto Mode became the default permission mode for Pro, Max, and Team plans on August 14, 2026, just days before these fixes shipped.

Claude Code

Claude Code Fixes a Memory Leak in Long Sessions With Subagents

Claude Code 2.1.238 fixes a memory leak in long interactive sessions where subagent tool results accumulated in memory indefinitely instead of being released once they scrolled out of view. Developers have filed a string of GitHub issues describing memory climbing into double-digit gigabytes during multi-hour sessions, particularly ones with heavy subagent use. The fix is notable because subagent forking became Claude Code's default behavior in version 2.1.232, which likely made this kind of memory growth more common for ordinary interactive sessions.

Claude Code

Claude Code VS Code Extension Adds Screen Reader Support

Claude Code's VS Code extension now supports screen readers in its transcript pane, shipped in version 2.1.236 on August 19, 2026. Live announcements now fire for replies, permission requests, errors, and status changes, and each turn gets a heading so screen reader users can navigate directly to it. The change brings the VS Code extension into parity with the accessibility support Claude Code's CLI has offered for months.

Claude Code

Claude Code Extends Its Weekly Usage Limit Boost Through August 31

Anthropic extended Claude Code's 50% increase to weekly usage limits through August 31, 2026, for Pro, Max, Team, and Enterprise plans. The boost, first introduced in May 2026, had already been pushed back at least twice (most recently to an August 19 cutoff) before this latest extension. Anthropic said it hopes to make the higher limits permanent but cited continued strong demand for its models as the reason capacity remains tight, promising to share more as plans develop.

Claude Code

Claude Code Adds a Concise Output Style

Anthropic shipped a new built-in "Concise" output style for Claude Code that has Claude lead with the result and skip preamble and narration, while still completing work just as thoroughly as before. Developers can turn it on from Output style in /config, or by setting "outputStyle": "Concise" in settings.json, joining the existing Default, Explanatory, and Learning styles. The same release also fixed a prompt-caching bug that affected sessions routed through an LLM gateway or a custom base URL.

Claude Code

Claude Code Patches a macOS Sandbox Wildcard Deny-Rule Bypass

Anthropic hardened Claude Code's macOS sandbox in version 2.1.236 by fixing how wildcard read-deny rules are enforced. Previously, a rule such as denying **/.env could be sidestepped inside an otherwise-allowed read region, and simply renaming a protected file could slip it past the wildcard match entirely. The fix makes wildcard deny rules take precedence within allowed read regions, extends protection to a matched directory's full contents, and closes the rename-based bypass, so files developers explicitly protect (credentials, secrets, .env files) stay protected.

Claude Code

Claude Code Fixes Permission Dialog Over-Grant Bugs

Claude Code 2.1.235 patched two permission-dialog bugs that could leave a user granting more access than intended. Pressing Shift+Tab while typing a comment inside a permission prompt was silently approving the pending edit and granting session-wide edit permission, instead of simply closing the comment field. Separately, Anthropic corrected cases where a permission dialog's displayed text and its "don't ask again" option did not accurately reflect the scope of the grant it produced, and the option is now withheld whenever a dialog cannot fully display what is being approved.

Claude Code

Claude Code Adds Optional Spellcheck for the Prompt Input

Anthropic added an optional spellcheck setting to Claude Code that underlines misspelled words in the prompt input as a developer types. The feature works by shelling out to whichever spell-checking engine, aspell, hunspell, or ispell, is already installed on the machine, rather than bundling its own dictionary. It ships off by default and can be turned on in configuration, giving users a lightweight way to catch typos before they land in a prompt sent to Claude.

Claude Code