Anthropic puts $35 million in Claude credits toward open-source security, and the week in Claude

PLUS: Anthropic says Claude designed protein binders against 14 of 15 targets, computer use leaves beta, and Cowork reaches mobile on every paid plan

By SamAug 23, 2026

Good to see you. Here's your week in Claude, from Aug 16 to Aug 22.

This week, Anthropic put $35 million in credits toward open-source security. Anthropic says Claude designed protein binders against 14 of 15 targets. And computer use left beta.

Plus 12 quick hits. Every claim has a link.

Anthropic brings Mythos 5 to more defenders, with a $35M fund

Anthropic is letting more security teams use what Claude Mythos 5 finds. It keeps limits on direct access to the model.

On Aug 21, Anthropic said Claude Security scans now run on Mythos 5. Claude Security is in public beta for Claude Enterprise customers. It scans code for flaws and suggests patches.

Per Anthropic, Mythos 5 scans are billed as standard token use under your plan. There's no separate add-on. A human must review and approve every patch.

A new Defender Advantage Fund will give $35 million in Claude credits. It goes to groups helping open-source maintainers secure their software. It starts with a small number of larger pilot grants.

Anthropic is also working with security partners to build Mythos 5 into their tools. Users get only specific outputs, like a list of suggested patches. They can't prompt the model to, say, build an exploit.

Anthropic says it's early in this work.

Anthropic's Cyber Verification Program gives vetted defenders reduced safeguards on Opus and Sonnet. In the coming weeks, it will add broader dual-use capabilities on those models. Mythos-class access is set to follow.

Anthropic says Claude designed protein binders against 14 of 15 targets

Claude designed small proteins that grab onto other proteins. Outside labs found they worked for 14 of 15 targets, per Anthropic.

The design work used Claude Mythos Preview and Opus 4.8, per Anthropic's Aug 18 post. It ran with minimal human involvement beyond the first prompt. Adaptyv Bio and Twist Bioscience made and tested the designs.

Per Anthropic, 22% to 35% of Claude's designs bound, depending on the setup. It says 10% to 15% is typical in protein design campaigns today. Some of its strongest designs bound several times more tightly than the best published result.

Anthropic plans more tests to confirm its hit rates and binding measurements.

In a second test, per Anthropic, Opus 5 worked through a contract lab's raw chemistry files. It returned results in 23 and 19 minutes. They matched the lab's own analysis on hydrogen counts and purity (96.4% versus 96.33%).

Protein design and other dual-use biology work aren't open for general access in Claude Fable 5.

Computer use, the Skills API and the Files API leave beta

Three agent tools on the Claude Platform are now generally available. There's also a new tool for working in web apps.

Computer use is out of beta on the Claude API, per the Aug 19 release notes. Claude can now take several actions per turn. Anthropic's Aug 20 post says tasks finish in fewer calls and less time.

The new browser use tool drives a browser that your app hosts. It reads the page itself, on top of screenshot-and-click control. Both tools work with Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8.

The Files API now has automatic file expiration. It also has 5x higher rate limits and 1 TB of storage per organization.

Computer use is now eligible for HIPAA-regulated workloads under Anthropic's BAA. Existing beta integrations keep working while you migrate.

Claude Code can now draft UI designs for you to pick from

Claude Code got a design tool, a shorter reply style and phone access to your machine.

It shipped versions 2.1.234 to 2.1.239 from Aug 17 to 21, per its weekly digest.

/design (research preview) brings Claude Design's artboards into the CLI and Claude Code Desktop. Claude publishes a canvas of editable artboards for your UI. You pick one, and Claude builds it.
/design is on Pro, Max, Team and Enterprise.
Concise is a new built-in output style. Claude leads with the result and skips preamble. Anthropic says Claude does the work as thoroughly as in the Default style.
Remote Control is out of research preview. Any machine running it now shows up as a device card in the Claude app's Code tab. Tap it to start a session there.
Claude Code now continues your session when a claude.ai usage limit resets. You can turn that off in /config.

Claude Cowork is now on mobile and web for all paid plans

Every paid Claude plan can now use Cowork on phones and on the web.

On Aug 18, Anthropic's Claude account posted the news. The mobile and web rollout began in July with the Max plan. That's per Anthropic's July 7 release note.

That note says Cowork runs your sessions remotely there (in beta). Your sessions and files are saved to your Claude account. Work continues when you close your laptop.

Quick hits

The 50% boost to weekly Claude Code limits now runs through Aug 31 (Aug 18). Anthropic's ClaudeDevs account says it hopes to make it permanent. It adds that capacity "may be tight over the coming weeks."
Claude had a service disruption on Aug 16, from 21:58 to 22:34 UTC (4:58 to 5:34 pm Central). Anthropic's status page first reported a sign-in problem. It covered claude.ai, platform.claude.com, the Claude API, Claude Code and Cowork.
The Python SDK reached v1.0 (Aug 20). It needs Python 3.10 or later. It removes long-deprecated features, including the temperature, top_p and top_k settings on Messages methods.
Claude Code 2.1.239 adds /claude-api upgrade (Aug 21). It moves Python projects from version 0.x of the SDK to 1.x.
Workbench is now playground in the Claude Console (Aug 18). It supports every Messages API parameter. It shows the full SDK request and the API response for each run.
Managed Agents builders can now limit which sites an agent can reach with web search and web fetch (Aug 19). That's per the release notes.
Admin API user-management endpoints for Claude Enterprise are out of beta (Aug 19). They cover members, invites, groups and custom roles.
Anthropic launched Claude Academy (Aug 20), with courses and tutorials on using AI. Some of its lessons aren't tied to any one product or model. It's at academy.claude.com.
Claude Code now tells Claude to use your email only to identify you (2.1.234, Aug 17). It's not to send it to unrelated services unless you ask.
Claude Code's built-in claude-api skill got lighter (2.1.234, Aug 17). Its context cost fell from ~200k+ tokens to ~25k. It now loads reference docs on demand.
Claude Code's macOS sandbox got stricter (2.1.236, Aug 19). Wildcard rules that block reading files like .env can't be bypassed by renaming the file.
MCP's maintainers published a new roadmap (Aug 22). Its priority areas include agentic messaging primitives, and agent identity and enterprise-ready security.

Worth your time

Claude on call (Aug 18): Claude Tag is the first responder for CI/CD failures at Anthropic. Claude wrote the first situation report in every recent incident that had one. It typically posted its first analysis within 15 minutes.
How monday.com rebuilt around Claude (Aug 20): monday rebuilt its platform around people and agents working together. Since the May 2026 launch, customers have had over 5 million agent interactions.
How ABC Legal turned every employee into a builder (Aug 17): the firm rolled out Claude Enterprise to 1,100 employees. Its CTO then moved their agents off personal desktops onto Claude Managed Agents.

That's the week. The next issue lands Sunday, Aug 30.

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

— Sam