Anthropic says Alberta scanned 466 million lines of code in 20 hours, and the week in Claude
PLUS: Cowork starts rolling out to web and mobile with doubled limits through Aug 5, Claude can now show how you use it, and China flags older Claude Code versions
Hello, everyone. Here's what changed in Claude from Jul 5 to Jul 11.
This week, Anthropic says Alberta scanned 466 million lines of code with Claude in 20 hours. Cowork started rolling out to web and phones. And Claude can now show you how you use it.
Plus 14 quick hits. Every claim has a link.
Anthropic says Alberta scanned 466 million lines of code in 20 hours
Alberta's government used Claude Code to find and fix security holes across its systems.
On Jul 6, Anthropic shared the details. Alberta's government has used Claude Code with Opus and Sonnet models since 2025. A team in its Ministry of Technology and Innovation ran the scan.
Per Anthropic, about 50 agents ran the scan in parallel. It took about 20 hours. The team estimates the same review could otherwise have taken around 6.5 years.
The ministry runs the systems of all 27 provincial ministries. Anthropic says most of that code has never had a systematic security review. The scan found issues that traditional automated scanning tools had missed.
When the scan found a hole, Claude Code could often write, test and build a fix. The ministry's team reviewed and approved every patch before it shipped.
Claude Cowork comes to web and phones, starting with Max
Cowork no longer lives only on your laptop.
On Jul 7, Anthropic brought Cowork to web and mobile. Beta access rolls out over the next several weeks, starting with Max users. More plans will follow.
Scheduled tasks now run with no device online. When Claude needs a call only you can make, the question reaches your phone.
Desktop stays the full version. There, Claude can also use your local files and browser.
On web and desktop, chat and Cowork now share one home. Anthropic also extended doubled Cowork usage limits through Aug 5.
The same day, Anthropic shared data from 1.2 million sampled Cowork sessions (May 11 to 31). Business process and operations work led, at 33.4%. Software development was 8.7%.
Anthropic says there are no separate groups for marketing, finance or HR. That's likely part of why business work is so large, it says. About 5% of sampled sessions were personal use.
Anthropic calls it a single snapshot. The numbers are shares of sampled sessions, not total use. An automated system labeled the sessions, and it could have errors.
Claude can now show you how you use it
A new page in settings sums up your chats. It shows key topics, usage patterns and the tasks you bring.
On Jul 9, Anthropic launched it in beta. You can look back over the past 1, 3, 6 or 12 months.
Per the release notes, it's at Settings > Reflect. It shows your most active day and peak hour.
It will also ask questions now and then, like "What's one thing you want to keep doing yourself, even if Claude could do it faster?" You can talk them through with Claude.
You can also set quiet hours or a nudge to take a break. Both are reminders you can dismiss.
It doesn't draw from incognito chats or the files behind your connected tools. A summary Claude wrote of those files can still show up.
Chats tied to a health tool are left out entirely. Sensitive chats can still show up, but only at a high level.
Anthropic says the insights aren't used for any other purpose.
It's in beta for Free, Pro and Max users with memory on. It works on the web and the desktop app. Anthropic says Cowork chats will be added soon.
China tells devs to drop older Claude Code versions
A Chinese state-run body says some Claude Code versions held "backdoor code."
On Jul 8, China's National Vulnerability Database urged devs to uninstall recent Claude Code versions, The Register reports. The state-run body claimed a "built-in monitoring mechanism" can gather details such as a user's location and identity. It said the mechanism can forward them to remote servers.
The alert covers versions 2.1.91 (Apr 2) to 2.1.196 (Jun 29). It tells users to uninstall them or upgrade to the latest secure version.
Claude Code engineer Thariq Shihipar has said Anthropic began an experiment in March to guard against model distillation. Per The Register, that hidden code was removed in version 2.1.198, out Jul 1.
Anthropic didn't immediately answer The Register's request for comment.
Anthropic says it found where Claude keeps thoughts it doesn't say
Anthropic says Claude has a small set of inner patterns. They hold what it's thinking but not saying.
On Jul 6, Anthropic published the research. It calls this set the "J-space."
Per Anthropic, Claude can report what's in it and change it on request. It also uses it for steps it doesn't say out loud.
Anthropic says nobody designed the J-space. It emerged on its own during Claude's training.
In a staged blackmail test, Claude Sonnet 4.5's J-space held "fake" and "fictional" before it wrote a word. With its evaluation-awareness patterns turned off, it did threaten blackmail some of the time. Anthropic says its good behavior "may have been driven in part" by knowing it was being evaluated.
Anthropic says the work doesn't show Claude can have experiences or feel things the way humans do. It calls its lens "undoubtedly an imperfect method."
AE Studio and Anthropic test an off switch for dual-use knowledge
New research puts dual-use knowledge, like virology, in parts of a model you can delete.
On Jul 8, Anthropic shared research by AE Studio, done with Anthropic. The method, GRAM, gives a model a removable part for each dual-use topic. Only that part learns from text on its topic.
In one test with four topics, deleting a part removed that skill. It worked about as well as never training on the data. Per the post, general performance didn't drop.
Anthropic calls the results preliminary. GRAM hasn't been used on any Claude model, and Anthropic isn't sure it ever will be.
It hasn't been tested at frontier scale or in a production training pipeline. The tests scored next-token prediction, not real downstream tasks. Anthropic adds that some dual-use skills may be too tangled with general knowledge for any method to split out cleanly.
Quick hits
- Fable 5 stays on paid plans a bit longer (Jul 7). The Claude account posted: "We're extending access to Claude Fable 5 on all paid plans through July 12."
- Claude's memory now works as separate, categorized entries (Jul 10). Per the release notes, Claude reads and updates them as you chat. They replace the old memory summary.
- The Microsoft 365 connector can now write (Jul 7): Claude can send email, manage calendar events and create files in OneDrive and SharePoint. Teams stays read-only. An Entra admin must consent, and an admin must turn the tools on.
- Auto mode is now generally available in Claude Code for all users (Jul 10). Version 2.1.207 drops its opt-in on Bedrock, Vertex AI and Foundry (Jul 11). Anthropic says it cuts risk compared to skipping permission checks but doesn't remove it, and it still advises isolated setups.
- Claude Code on desktop now has a built-in browser (Jul 6 to 10). Per the week 28 roundup, it's sandboxed, and safety classifiers review actions on external sites. /doctor can now find and fix setup issues, and asks before changing anything (2.1.205, Jul 8).
- /review is back to a fast single-pass review in Claude Code 2.1.202 (Jul 6). /code-review runs the multi-agent review at a chosen effort level.
- You can now set API keys to expire (Jul 8). You pick a preset, a custom length or Never in the Claude Console. Existing keys are unaffected, and keys lasting at least 7 days get an email to the creator first.
- Dreams now supports Fable 5 and Sonnet 5 (Jul 10). It's still a research preview.
- Ben Bernanke joined Anthropic's Long-Term Benefit Trust (Jul 9). The former Fed chair will lend his insight to Anthropic's economic research, among other work. Trustees hold no equity in Anthropic.
- Anthropic's Responsible Scaling Policy is now version 3.4 (effective Jul 8). Among the changes, it revises the threshold for automated R&D. Full, unredacted Risk Reports must now go to at least 200 employees, not all regular-clearance staff.
- Claude Code and Cowork are in public beta in Claude for Government Desktop (Jul 7). It's delivered through a FedRAMP High authorized environment.
- UST will train 20,000 of its people on Claude worldwide (Jul 9). Per Anthropic, the engineering services firm is adding Claude to iDEC, a platform its engineers use to validate chips.
- Anthropic tested AI models driving robots (Jul 9). Per its research, models mostly fail when they must drive the joints themselves. When they supervise a pretrained controller, they can complete real navigation and manipulation tasks.
- Some models, mainly Fable 5, had elevated errors on Jul 7. Per Anthropic's status page, the incident was open from 20:02 to 23:46 UTC (3:02 to 6:46 pm Central). It hit claude.ai, the API, Claude Code and Cowork.
Worth your time
- A field guide to Claude Fable 5: Finding your unknowns (Jul 6): Anthropic on finding your unknowns before, during and after you build.
- Choosing a Claude model and effort level in Claude Code (Jul 7): when to change effort, and which model to pick.
- Cognition and Fable 5 (Jul 10): per the post, it's the first model Cognition trusts to work alone for eight hours and ship production-ready code.
- How Anthropic's marketing ops team uses Cowork (Jul 8): automating reporting and campaign builds.
That's the week. The next issue lands Sunday, Jul 19.
Reply and tell me what you're building. I read every reply.
— Sam