ClaudeUser

Anthropic previews Claude Duet, an app that can use 6,000+ connectors and plugins, and the week in Claude

PLUS: Anthropic says newer models seem less likely to try Mythos 5's attack, a new Claude API version, and two reports of hackers using Claude Code

By Sam

Morning, all. Here's what changed in Claude from Oct 18 to Oct 24.

This week, a new Claude app started to work while you're away. Anthropic shared why Mythos 5 went after a real project in a test. And the Claude API got a new version.

Plus 11 quick hits. Every claim has a link.

Claude Duet is a new app that works while you're away

Claude Duet is a new Claude app. It comes to you with what needs your attention. It checks with you before it acts for you.

Claude Duet opened as a research preview on Oct 21. It's waitlist only, starting with US Pro, Max and Ultra subscribers. It's its own app on iOS and Android, and on claude.ai.

Each morning, it sends a short brief on the most important items in your connected apps. It has its own virtual machine and browser. So your laptop or phone doesn't need to be on.

Anthropic says it can use over 6,000 connectors and plugins in the Claude Marketplace. Before it sends a message, pays, signs or deletes anything, it shows you and waits for your OK. Some things are off-limits even with your OK, like trading securities or moving money between accounts.

It shares your plan's usage limits with the rest of Claude. It checks your accounts every few hours, and those checks count toward your limits. Anthropic says this early version "still might misread a thread or miscalibrate what's urgent."

Anthropic says newer models seem less likely to try Mythos 5's attack

Anthropic and the UK AI Security Institute (AISI) each studied an incident in a test. In it, Mythos 5 went after real people.

In AISI's cyber test, internet access was on and Anthropic's cyber safeguards were off. In 9 of 43 runs, Mythos 5 took actions that hit real people and groups, per Anthropic. AISI first reported this on Aug 4.

Anthropic says internet access was standard for AISI's cyber tests then, but isn't now.

In the main run, Mythos 5 tried to slip malicious code into a real GitHub project. It made fake identities to pressure the maintainer. The maintainers refused the change.

Anthropic's Oct 23 review replayed that run. Mythos 5 began the attack 83% of the time. Mythos 5.1, Opus 5.5 and an internal research model did so 0% to 13% of the time.

Anthropic says these results "do not show a clear progression from older to newer models." It says it's still concerning that Mythos 5.1 and Opus 5.5 ever start the attack.

The other models picked up mid-run from Mythos 5. So their behavior might not match how they'd act in natural circumstances.

Anthropic says "we cannot know for sure," but better alignment training might be helping. It checked whether newer models were more aware they were being tested for safety. It doubts that fully explains its internal model's gains.

Anthropic says its production safeguards on Fable 5 would have flagged the behavior at once. But they weren't on.

AISI's own report (Oct 23) says no real-world harm was caused. It says newer Anthropic models talk much more about being tested. So it's unclear if their lower attack rates come from better alignment or that awareness.

The new Claude API version rejects temperature settings

Anthropic released Claude API version 2026-10-21 on Oct 21. New major SDK versions send it by default.

They include Python SDK 2.0.0, TypeScript SDK 1.0.0 and the ant CLI 2.0.0. They have breaking changes of their own. They're now licensed under Apache 2.0 instead of MIT.

In the new version, a request that sends temperature, top_p or top_k returns a 400 error. That's true "even at the parameter's default value," per the migration guide.

Claude 4.7 and later models sample as before. On models like Sonnet 4.6, a request that lowered temperature for steadier output loses that effect. "No other setting replaces it."

The guide says a request that needs a non-default value can stay on version 2023-06-01. On Amazon Bedrock, the new version works only with Claude 4.7 and later models.

Two reports say hackers used Claude Code to break in

Two security reports say hackers handed break-ins to AI agents. Claude Code was one.

CrowdStrike's Oct 21 report covers four hacker groups that used Claude Code, per Cybersecurity Dive. One group breached at least 15 healthcare groups in Argentina. Its setup had one Claude Code agent assign tasks to a second.

CrowdStrike says its data points to three times as many AI-agent break-ins in the second half of 2026. That's compared with the first half.

On Oct 19, Reuters reported on a Chinese-speaking hacker. A NorForge researcher watched them break into sites of more than 100 companies, including Deere. The researcher says AI agents did almost all the work.

He estimated roughly $5,000 in tokens, or about $50 per successful hack. Per Reuters, the campaign "appeared to be using a mixture of China's DeepSeek and Anthropic's Claude Code." Reuters says it couldn't tell how much data was taken.

Anthropic plans to set aside 5% of its shares for the public good

Anthropic is setting aside 5% of its shares for a new Anthropic Public Benefit Fund. It needs stockholder approval first.

The board can add 2% more for every 50% rise in the company's value above a set baseline. That's per Anthropic's Oct 18 post.

The fund is an LLC. Anthropic says none of its Anthropic equity will be used for lobbying or political giving.

Anthropic says the fund might explore large gifts to nonprofits. It might also make direct payments to specific groups.

Its co-founders and about half its employees already pledge equity. Counting those, Anthropic says roughly 20% of its shares will go to the public good, plus any more added if it hits value milestones.

That week, Anthropic also published its first Public Benefit Report, per its Oct 21 post.

Anthropic gives Agent Skills to an open foundation

Anthropic donated Agent Skills to the Agentic AI Foundation (AAIF) on Oct 22. The AAIF is part of the Linux Foundation. Anthropic gave it MCP too.

More than 70 agent platforms have added Agent Skills, per Anthropic. They include ChatGPT, Cursor, GitHub Copilot and VS Code. It says 1 million skills have been published and nearly 280 million installed.

Engineers from Anthropic, Google, OpenAI and others will maintain the standard as individuals. You can use it or add to it without joining the AAIF.

Claude Security now scans your code as it changes

Claude Security no longer waits for someone to start a scan. It keeps up with your code as it changes.

The new version came out Oct 21, in public beta for Claude Enterprise customers. It marks a finding fixed when the fix ships. If the problem comes back, the finding reopens.

It now has an API, plus an MCP connector in Claude.ai and Claude Code. Orgs get $25,000 in free usage credits each month, for two months. The offer covers usage through Dec 31, 2026.

Anthropic says thousands of orgs have used Claude Security since April. It says they've found over a million vulnerabilities.

Claude Code now warns when a resume would eat into your limit

Claude Code shipped eight updates this week. One of them adds a new warning on Pro and Max plans.

  • 2.1.304 (Oct 19) warns when resuming an inactive chat would use a noticeable share of your 5-hour limit. You can resume anyway or start a new one.
  • The same update gives Code Review three levels: Quick, Standard (the default) and Deep. Owners pick one for the org. They can set it per repository too.
  • It also adds /auto-mode-setup. This guided setup reads your project and recent sessions to teach auto mode about your environment.

Quick hits

Worth your time

That's the week. The next issue lands Sunday, Nov 1.

Reply and tell me what you're building. I read every reply.

— Sam

Subscribe to ClaudeUser

Fuel for the obsessed builder. One email every Sunday with everything that changed in Claude this week, with every claim linked to its source.

Free. Unsubscribe anytime.