Anthropic says Claude leads 26% of its AI research work, and the week in Claude

PLUS: a 3-person team used Claude to break into OpenAI, chat and Cowork merge into one Claude, and Claude Code projects

By SamSep 20, 2026

Hello, builders. Here's what changed in Claude from Sep 13 to Sep 19.

This week, Anthropic says Claude leads 26% of its AI research work. Chat and Cowork started merging into one Claude. And a small team used Claude to break into OpenAI.

Plus 10 quick hits. Every claim has a link.

Anthropic says Claude now leads 26% of its AI research work

Anthropic shared new numbers on how much of its own AI work Claude now does.

Anthropic's Sep 17 post says that as of August 2026, Claude "leads" 26% of its AI R&D work. That means Claude can do most of a task from a high-level prompt. A human supervises.

Per Anthropic, over 90% of the work is at or above the level where AI "collaborates." Claude isn't fully autonomous for any part it measured.

Claude agents gathered the evidence. A Claude judge rated it. Anthropic says the judge model "could make the same kinds of errors as the model it is checking."

In August, a monitor blocked about 1 in 47,000 of over a billion decisions by research and engineering agents. That's on Anthropic's most-used internal platform, and Anthropic says the figures cover that platform only.

Claude chat and Cowork are now one Claude

Anthropic is merging chat and Cowork into one Claude. Claude works out what each task needs.

Anthropic says the merge started on Sep 16. It's rolling out on Pro, Max and Ultra plans over the next few weeks. Team and Free plans "will follow soon."

If you've been using Cowork, your chats, projects, artifacts, connectors and skills are where you left them.

Claude Docs and Claude Slides are new. Claude Design now works inside your chats too.

All three are in beta on paid plans. Enterprise admins choose when to turn them on.

By default, Claude asks before it takes an action. You can let it keep working instead. Then it checks in only when something needs a closer look.

A 3-person team used Claude to break into OpenAI staff accounts

Security startup Hacktron says it used Claude to take over several OpenAI staff accounts. It reported the flaws through OpenAI's bug bounty program.

In a Sep 13 post, Hacktron says the way in was an image upload bug on OpenAI's forum. From there, it reached OpenAI staff ChatGPT accounts. It says those could reach OpenAI's internal code.

Hacktron says Opus 4.8 "struggled across several sessions" to build a working exploit with ASLR on. "Within hours of Opus 5's release," Opus 5 did it.

Hacktron says Opus refused to write exploits for remote servers. So the team made its own server look like a hacking contest target.

OpenAI confirmed its fix about 14 hours after the report. It paid a $6,500 bounty.

OpenAI says testing against its forum was excluded from its bug bounty. It says the award covers its own flaw, not the attack on the forum.

Hacktron says its work was "not completely autonomous hacking." Skilled humans still guided it.

Claude Code projects can now split one goal across many threads

In a Claude Code project, you now set a goal and Claude runs the work.

Per Anthropic (Sep 17), Claude scopes the request.

It hands parts to threads that run at the same time. Then it reviews and joins the results. Each thread is a Claude Code cloud session on its own branch.

Threads share a memory, so the project builds context over time. Anthropic says projects can hit usage limits faster. Each thread is a full Claude Code session.

It's in beta for select Pro and Max users who use cloud sessions in Claude Code. They also must have no existing projects on the web or desktop. More Claude Code users on those plans get it over the coming week, per Anthropic.

Accenture will test Anthropic's models from the inside

Anthropic picked Accenture to check its AI from inside the company.

The Sep 18 deal is led by Faculty, Accenture's AI business. It covers red-teaming models, alignment checks and testing safeguards.

Anthropic and Accenture each expect to invest at least $1 billion in this area. That's over the next five years.

Per Anthropic, these evaluators will get access close to an employee's. It says many details "are still being worked out." No standards exist yet for what they can see or report.

Anthropic will pay for Accenture's work directly. The deal isn't exclusive. Anthropic says it will name more evaluators in the coming weeks.

Vetted life science teams can get Claude with fewer biology blocks

Anthropic opened a program that loosens Claude's biology safeguards for vetted teams.

The Life Sciences Verification Program launched in beta on Sep 17. It gives life science pros Mythos, Opus and Sonnet models. Their safeguards are "more permissive for biology-related work."

Teams first pass checks of their research credentials, security and ethics oversight.

For now, it works with the API in Anthropic's own console. It also works on Claude Enterprise and Team plans.

It doesn't cover individual plans or third-party platforms yet. During the beta, it isn't open to orgs with a BAA. All other safeguards, like cyber classifiers, stay in place.

Anthropic keeps program traffic for 30 days to watch for misuse. It says that data can't be used to train its models.

Claude Code now reads AGENTS.md files

Claude Code shipped eight updates this week. One lets it read AGENTS.md instruction files.

2.1.277 (Sep 18) reads AGENTS.md in a project with no CLAUDE.md. You can change this in /config.
It also marks subagent results as subagent output. Anthropic says text in them can't pass as the session's own instructions.
2.1.278 (Sep 19) gives Claude API and Enterprise users a server-side auto mode checker by default. It doesn't charge for the checker's overhead. Claude Code warns if it falls back to a billed check.
2.1.275 (Sep 17) syncs the skills and plugins you turned on at claude.ai. They reach terminal sessions signed in with that account. You can opt out.
2.1.275 also stops running install scripts for plugins installed from npm.

Quick hits

Salesforce in Claude is now in beta (Sep 15). The plugin brings a seller's accounts, opportunities and pipeline into Claude, under their existing Salesforce permissions. Per Anthropic's release notes, it's on all paid plans for orgs Salesforce approves through its beta sign-up.
Claude for Small Business now has 43 workflows and 27 new integrations, like Shopify, Stripe and Zapier (Sep 15). Anthropic says it has been installed more than 900,000 times.
AIR Security disclosed "Plugin4Shell" (Sep 17), a zero-click flaw it found in how Claude Code, Codex, GitHub Copilot and Gemini CLI install plugins. For Claude Code, AIR says it works only where a branch can be named like a commit hash, which GitHub rejects. AIR says Anthropic patched Claude Code in 2.1.179.
The Messages API can now compact a chat on demand on the Claude API (beta, Sep 14). It returns a summary block. You send it in place of the old messages.
The Compliance API now returns Claude in Chrome transcripts (beta, Claude Enterprise, Sep 18), per the release notes.
Claude Code 2.1.271 (Sep 14) brings fast mode to Remote sessions, where your org allows it.
Claude Code 2.1.275 added a send-now key (Sep 17). It stops the current turn and sends all your queued messages at once.
OpenAI says it has worked with Anthropic and Google DeepMind on AI safety for weeks, TechCrunch reports (Sep 15), citing Bloomberg. OpenAI's policy chief said so on Tuesday.
Anthropic confirmed it runs a wet biology lab in the Bay Area, TechCrunch reports (Sep 18). Anthropic gave no specifics but said its main focus is basic biology, not drug discovery.
Anthropic joined a grid group formed by Emerald AI, Google and Nvidia (Sep 17). Per TechCrunch, the group says data centers could cut power use when the grid is busy. It says that could let another 100 gigawatts of data centers connect.

Worth your time

How Claude sped up biology models (Sep 17): Anthropic says Claude, with two staff supervising, made over 30 open-source biology models about 4x faster on average. Precision dropped slightly. It took under four weeks.
Agentic coding is straining CI (Sep 14): an Anthropic engineer says its CI jobs grew 25x in six months. The post explains how the team rebuilt its test selection service.

That's the week. The next issue lands Sunday, Sep 27.

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

— Sam