Claude in Chrome opens to everyone, and agents get hands on lab hardware

TLDR;

Anthropic previewed the Model Hardware Standard, a model-agnostic spec that lets AI agents operate physical lab and manufacturing equipment in parallel — a significant platform expansion for builders in science and robotics. Claude Code v2.1.251 shipped five security fixes that patch real path-traversal and permission-bypass vectors, so update immediately if you run Claude Code in production. Claude in Chrome exited pilot and is now generally available on every paid plan, with autonomous browser actions gated by a safety classifier. Model retirements for Sonnet 4.5, Haiku 4.5, and Opus 4.5 are now inside the planning horizon — deprecation dates below.

Headline

Model Hardware Standard — Research Preview

Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification that lets AI agents safely operate physical lab and manufacturing devices — microscopes, liquid handlers, robotic arms, laser rigs — in parallel. MHS is model-agnostic (not Claude-only) and is initially available to select science, robotics, and manufacturing organizations, with plans to eventually open-source the standard so any device manufacturer can adopt it. If you build in scientific computing, lab automation, or industrial robotics, this is the most consequential platform signal of the week.

1. API & Model Updates

Claude Sonnet 4.5 / Haiku 4.5 / Opus 4.5 — Retirement Dates

Model retirements are now close enough to require sprint-level planning: Sonnet 4.5 retires September 29, Haiku 4.5 retires October 15, and Opus 4.5 retires November 24, 2026. Any production integration pinned to these model IDs needs a migration path now.

2. Skill of the Week

/break and /variant, two new interface skills from Jakub Krehel

This week's editor pick. Jakub Krehel (@jakubkrehel) added two new skills to his open-source interface skills collection (Aug 25), which now sits at 4,600+ GitHub stars. /break renders a component you choose in every state and worst-case scenario on a temporary page and stress-tests it — a fast way to find design-system issues you did not know you had. /variant builds three deliberately different versions of a component behind a picker on the real page, so you flip between them and choose, instead of accepting the first design Claude produces. Both slot into a collection of eleven skills covering accessibility, color, layout, and typography. → GitHub repo

3. Tool of the Week

Ultramock

This week's tool pick. Ultramock, the browser-based tool by Josh Millgate (@joshmillgate) that turns product screens and screen recordings into premium visuals — device mockups, multi-video timelines, animated renders — shipped a new Lens blur feature (Aug 27) that adds realistic depth-of-field to renders. If you ship product visuals for launches, social posts, or your landing page, this replaces an After Effects workflow with a browser tab. Used for the render work behind t3 code's theme launch, among others.

4. Community Releases

Archify (@aitech_komoriya highlight) — An agent skill that turns a codebase or description into interactive architecture, workflow, sequence, data-flow, and lifecycle diagrams as a single self-contained HTML file, with PNG/SVG/WebM export. It builds a typed JSON intermediate representation and verifies structure before drawing, so it does not invent plausible-looking connections. Works with Claude Code, Cursor, Codex, and OpenCode; the repo sits at 36,000+ GitHub stars. → X post · GitHub

genart-skill (@camillerouxart) — A free, open-source Claude Code plugin that teaches Claude the craft of generative art: deterministic output from a hash, resolution independence (the same piece at 400px and 4000px), honest rarity distribution, and tools to verify a sketch before minting. → X announcement · GitHub

x64dbg MCP Server (@duty1g) — A native MCP plugin for x64dbg, the classic open-source Windows debugger, that hands debugging capability directly to Claude Code, Codex, and other agents: AI-assisted breakpoints, memory inspection, and reverse-engineering workflows. Released Aug 22 and already past 1,700 GitHub stars after going viral in the reverse-engineering community. → GitHub

5. Claude Code & CLI

SendFeedback Tool & `/claude-api cost-optimize`

Two quality-of-life additions landed this week: the SendFeedback tool lets Claude draft a feedback report during a broken session for you to review and send from /feedback. The /claude-api cost-optimize command profiles an existing project's API spend and walks through cost levers — caching, token hygiene, batching, effort settings, model choice — one measured change at a time.

`--restricted` Mode & Per-Agent Cache TTL

claude --restricted (or CLAUDE_CODE_RESTRICTED=1) removes command-execution and WebFetch tools, keeps file access inside the working directory, refuses bypassPermissions, and ignores user/project settings files — useful for building tightly sandboxed Claude Code integrations. Also: experimental.cacheTtl in agent frontmatter sets a per-agent prompt cache TTL ("5m" or "1h").

Spend Limit Bar & Prompt-Cache Visibility

/usage now shows a spend limit bar for sessions behind a Claude Apps gateway. /cost adds a per-session prompt-cache line (hit ratio, misses, tokens re-cached, warm/cold) and a matching prompt_cache object for status-line scripts — giving builders better visibility into cache efficiency without external tooling.

6. MCP Ecosystem

Enterprise-Managed Authorization for MCP Connectors — Generally Available

Enterprise-managed authorization for MCP connectors reached general availability on Aug 24, expanding to Datadog, Notion, and Slack (joining Asana, Atlassian, Canva, Figma, Granola, Linear, and Supabase). Admins on Team and Enterprise plans provision connector access centrally via their identity provider; users get tools connected without individual OAuth flows. Built on the Identity Assertion JWT Authorization Grant (ID-JAG). Exa, Miro, and Zoom are listed as coming soon.

MCP Transports Working Group Chartered

The MCP Transports Working Group was chartered on Aug 26, targeting a 2026-12-15 specification release. Builders maintaining custom MCP transports should track this working group — charter and participation details are on the MCP roadmap page.

Per-Server MCP Timeout in Agent SDK TypeScript

createSdkMcpServer({ timeout }) now accepts a per-server timeout that overrides the global MCP_TOOL_TIMEOUT for that server's tool calls — useful for MCP servers with high-latency upstream APIs.

7. Security

Claude Code v2.1.251 — Five Security Fixes, Update Immediately

v2.1.251 patches five exploitable issues: (1) symlink swap — file tools (Read, Write, Edit) followed a symlink swapped inside the working directory after the permission check, allowing reads/writes outside the approved location; (2) plugin path traversal — plugin commands could point outside the plugin directory; (3) beta tracing bypass — project settings could enable detailed beta tracing or raw API body logging, and a lower-scope endpoint could bypass an OTLP collector pinned by managed settings; (4) workflow tool path leak — the Workflow tool could read a scriptPath outside what the session permitted; (5) Windows NT device-prefix bypass — \\?\ paths bypassed UNC path validation, closing an NTLM credential-leak vector. Update to v2.1.251 immediately if running Claude Code in any production or shared environment.

GitLab Fixes Claude AI Agent Flaw — CVE-2026-18252 (CVSS 7.3)

GitLab patched a high-severity flaw in its Duo Claude AI agent (Aug 26) that allowed authenticated developers to execute arbitrary commands within CI pipeline contexts. Affects GitLab EE 18.9 through 19.3.1. Self-managed GitLab customers must update to 19.3.1, 19.2.5, or 19.1.7 immediately. GitLab.com is already patched; Dedicated customers need no action.

Anthropic CVD Dashboard — 2,300 Vulnerabilities Disclosed

As of Aug 26, Anthropic's coordinated vulnerability disclosure program (using Claude Mythos Preview on open-source software) has disclosed 2,300 vulnerabilities across 392 projects, with 421 patched. The public dashboard now includes a severity-agreement matrix comparing Claude's assessments against external security firms. Relevant for builders who depend on open-source infrastructure.

8. Competitor Dev Tools

Cursor — "Start from Scratch, Without a Repo" (Aug 27)

Cursor Cloud Agents no longer require a connected SCM provider to start. Developers can prompt immediately, then save to a Cursor Origin repo automatically created in the background. The cloud agent's live environment is now port-forwarded to the browser for live preview and design mode, with a Vercel publish option. Builders evaluating cloud agent architectures should note this as a UX pattern worth watching.

OpenAI Codex v0.151.0 — MCP Tool Result Interception & Plugin Catalogs (Aug 29)

Codex v0.151.0 adds: extensions that can inspect or replace MCP tool results before they reach the model; combined per-repository plugin catalog configuration; and a configurable grace period for discovering tools from optional MCP servers. Builders maintaining cross-agent compatibility with Codex should review the plugin catalog changes.

9. Claude Apps

Claude in Chrome is Generally Available

Claude in Chrome exited its pilot and is now generally available on every paid Claude plan (Aug 26). Claude can now take actions in the browser autonomously — reading and typing text, clicking links, navigating pages, filling forms with your existing logins — with a safety classifier validating each action before it runs, instead of requiring per-action approval. Anthropic details its prompt-injection defenses in the post. If your product assumes users manually bridge Claude and web tools, this changes the baseline.

Claude Cowork Gets a Built-In Browser

Claude Cowork now ships with a built-in browser (Aug 26) — nothing to install. Cowork tasks that need web access run against a browser Claude controls directly, extending the agentic browsing surface beyond the Chrome extension.

Claude's Memory Works Everywhere

Chat and Claude Cowork now share one memory (Aug 25): context built up across months of conversations is available the moment you hand Cowork a task, and vice versa. Users can view, edit, or delete everything Claude remembers, topic by topic; sensitive subjects (health, beliefs) stay out of memory by default with an opt-in toggle.

10. Enterprise & Business

Expanding Support for Scientists

Anthropic announced expanded support programs for scientific researchers (Aug 27), including access to Claude for research workflows and the Model Hardware Standard preview. For builders in life sciences, materials science, or lab automation, this represents a new partnership pathway with Anthropic directly.

11. Research & Safety

Enabling Independent Research on How People Use Claude

Anthropic published a framework (Aug 26) for granting vetted external researchers access to anonymized Claude usage data. For builders: this signals Anthropic is moving toward more structured external transparency, which may affect how product usage data is handled under enterprise agreements.

Automated Researchers Can Reliably Mitigate Alignment Failures

Anthropic published research (Aug 28) showing that automated research agents can detect and mitigate certain alignment failures in Claude — a result with direct implications for how builders design oversight mechanisms in agentic pipelines. If you're building multi-agent systems, this paper is worth reading for its framing of automated monitoring as a safety layer.

Release Notes

Claude Code:

v2.1.251 (Aug 28) — PreModelSwitch/PostModelSwitch hook events; live subagent tool-call streaming to Remote Control clients; spend limit bar in /usage; per-session prompt-cache line in /cost; five security fixes (symlink swap, plugin path traversal, beta tracing bypass, workflow path leak, Windows NT prefix bypass) — update immediately
v2.1.250 (Aug 28) — Bug fixes and reliability improvements
v2.1.248 (Aug 27) — --restricted / CLAUDE_CODE_RESTRICTED=1 mode; experimental.cacheTtl agent frontmatter; self-hosted runner client label override; cross-session messaging on Bedrock/Vertex/Foundry; prompt-cache miss fix for long sessions
v2.1.247 (Aug 26) — SendFeedback tool; spinnerTipsOverride org customization; /claude-api cost-optimize command; Admin API coverage in /claude-api skill; keyboard shortcut fixes for non-Latin layouts
v2.1.246 (Aug 25) — Auto mode tab in /permissions; turn completion time in duration line; fullscreen/transcript rendering fixes; background session reliability fixes; MCP tool call interrupt fix in headless sessions
v2.1.245 (Aug 25) — Fixed crash on startup on glibc 2.44 systems (Arch Linux, CachyOS, Fedora Rawhide)

Agent SDKs (TypeScript):

v0.3.251 (Aug 28) — Parity with Claude Code v2.1.251
v0.3.250 (Aug 28) — Parity with Claude Code v2.1.250
v0.3.248 (Aug 27) — Per-server timeout for SDK-hosted MCP servers (createSdkMcpServer({ timeout }))
v0.3.247 (Aug 26) — Optional ambient flag on task entries; permissionMode live-value fix on system/init frames
v0.3.246 (Aug 25) — user_message_uuid on error/reply messages; modelUsage[*].costBasis field; modelPricing in managedSettings; perTaskStopAffordance option
v0.3.245 (Aug 25) — Parity with Claude Code v2.1.245

Agent SDKs (Python):

v0.2.148 (Aug 28) — Bundles CLI v2.1.251 (pip install claude-agent-sdk==0.2.148)
v0.2.147 (Aug 28) — Bundles CLI v2.1.250
v0.2.146 (Aug 27) — Bundles CLI v2.1.248
v0.2.145 (Aug 27) — Bundles CLI v2.1.247
v0.2.144 (Aug 25) — Bundles CLI v2.1.246

Action Items

Immediate:

Update Claude Code to v2.1.251 — patches five exploitable security bypasses (symlink swap, plugin path traversal, beta tracing bypass, workflow path leak, Windows NT device-prefix NTLM leak). Any production or shared-environment deployment is at risk on earlier versions.
Update GitLab self-managed to 19.3.1, 19.2.5, or 19.1.7 — CVE-2026-18252 (CVSS 7.3) allows arbitrary CI command execution via the Duo Claude agent. GitLab.com is already patched.
Update to claude-agent-sdk v0.3.251 (TS) or v0.2.148 (Python) to pick up the v2.1.251 security fixes in embedded CLI.

By September 29, 2026:

Migrate any production integrations pinned to Claude Sonnet 4.5 — retires September 29.

By October 15, 2026:

Migrate any production integrations pinned to Claude Haiku 4.5 — retires October 15.

By November 24, 2026:

Migrate any production integrations pinned to Claude Opus 4.5 — retires November 24.

By December 15, 2026:

Track the MCP Transports Working Group spec release — custom transport maintainers should review the draft spec when it drops to assess breaking changes.

Want to catch up? Read every past issue of This week in Claude.

Did you like this issue?

One tap is the fastest way to shape next week.

👍 Liked it · 👎 Didn't like it

And I read every reply. Hit reply and tell me what you want more of, or less of, and it goes straight into the next issue.