Anthropic Claude wants to be your helpful colleague, always looking over your shoulder
- Reference: 1768337791
- News link: https://www.theregister.co.uk/2026/01/13/anthropic_previews_claude_cowork_for/
- Source link:
The company said Cowork was inspired by Claude Code, its tool for generating programming code using natural language prompts and whatever files and tools have been made accessible.
With Claude Code, the output is generally source code or related artifacts like documentation or configuration files. With Cowork, the output may be local files like spreadsheets or interaction with desktop or web applications.
[1]
Cowork may find fans among knowledge workers who have to create presentations but take no pride in their work, who need to move data between applications without codified import and export rituals, or who have to conduct extensive reformatting operations in spreadsheets.
[2]
[3]
It may also find favor as a crutch that allows people with minimal application skills to delegate activities beyond their competency to an AI model with a different set of limitations. Think of all the charts Cowork can produce for those untutored in data science and statistics, or the proposals and position papers it can generate for summarization by other AI models.
But to function, Cowork must have access to files and applications.
[4]
That's where [5]the warning comes in. Anthropic makes it clear that Cowork is a research preview and should be used with caution. Users are advised not to give Cowork access to local files containing sensitive information.
When used in conjunction with the Claude in Chrome extension (which comes with [6]its own safety warning ), Cowork should be limited to accessing trusted websites. The same goes for modifying Claude's default internet settings.
Also, users are advised to "monitor Claude for suspicious actions that may indicate prompt injection." Yes, if you happen to sic the AI model on a file or website that contains text that could be interpreted as a system instruction rather than data, you could trigger an indirect prompt injection attack. Security researchers have [7]repeatedly [8]shown how [9]relatively easy it is to craft such attacks.
[10]
Anthropic goes on to offer reassurance that it implements various layers of protection, like reinforcement learning to convince Claude to refuse malicious instructions and content classifiers that attempt to catch malicious consent before it can be passed to the underlying model.
But the company follows its mollification with a cautionary reprise: " Important: While we've enacted these safety measures to reduce risks, the chances of an attack are still non-zero. Always exercise caution when using Cowork." You've been warned.
For those still not convinced to proceed carefully, there's a disclaimer of liability that insists "You remain responsible for all actions taken by Claude performed on your behalf."
In case that's not clear, the company says this includes content published or messages sent, purchases or financial transactions, data accessed or modified, and the terms of service at third-party websites, which, [11]like Amazon , may prohibit automated interaction without permission.
[12]AI and automation could erase 10.4 million US roles by 2030
[13]Developer writes script to throw AI out of Windows
[14]Trump says Americans shouldn't 'pick up the tab' for AI datacenter grid upgrades
[15]Linus Torvalds tries vibe coding, world still intact somehow
Once you've gotten past the disclaimers and the necessary permissions have been granted, and if you happen to meet the access requirements – a Claude Max subscription ($100 or $200 per month) and the [16]Claude macOS desktop app – then you can tell Cowork " [17]Please help organize my desktop ." With any luck, the assistive software will rearrange your messy desktop files more thoroughly than your Mac's Clean Up Selection command.
That's not all. "Claude can then read, edit, or create files in that folder," Anthropic explains in its [18]blog post . "It can, for example, re-organize your downloads by sorting and renaming each file, create a new spreadsheet with a list of expenses from a pile of screenshots, or produce a first draft of a report from your scattered notes."
Cowork can be extended to work with third-party apps and sites via [19]Connectors , and with specific file formats like PowerPoint (pptx), Excel (xlsx), Word (docx), and PDF (pdf) via [20]Skills .
The challenge is figuring out what makes sense to automate. Some batch operations on desktop files may be more efficient as CLI commands or shell scripts, which Claude might even compose if asked.
Anthropic has acknowledged that users may have trouble coming up with ways to employ Claude by publishing a list of [21]suggested use cases . These include tasks like " [22]clean up promotional emails ," " [23]organize files in Google Drive ," and " [24]thoughtful gift giving with Claude " .
Anthropic said it expects to make improvements in Cowork based on feedback, and to bring the automation service to Windows. ®
Get our [25]Tech Resources
[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aWlxojTVGpasd3I8RggnxAAAAso&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aWlxojTVGpasd3I8RggnxAAAAso&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aWlxojTVGpasd3I8RggnxAAAAso&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aWlxojTVGpasd3I8RggnxAAAAso&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://support.claude.com/en/articles/13364135-using-cowork-safely
[6] https://support.claude.com/en/articles/12902428-using-claude-in-chrome-safely#h_044f6a88a7
[7] https://www.theregister.com/2026/01/08/openai_chatgpt_prompt_injection/
[8] https://www.theregister.com/2025/08/08/infosec_hounds_spot_prompt_injection/
[9] https://www.theregister.com/2025/10/28/ai_browsers_prompt_injection/
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aWlxojTVGpasd3I8RggnxAAAAso&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://www.theregister.com/2025/11/05/amazon_perplexity_comet_legal_threat/
[12] https://www.theregister.com/2026/01/13/ai_us_jobs_2030/
[13] https://www.theregister.com/2026/01/13/script_removes_ai_from_windows/
[14] https://www.theregister.com/2026/01/13/trump_datacenter_power_costs/
[15] https://www.theregister.com/2026/01/13/linus_torvalds_vibe_coding/
[16] https://claude.com/download
[17] https://www.youtube.com/watch?v=WBNZpAWhw5E&t=1s
[18] https://claude.com/blog/cowork-research-preview
[19] https://claude.com/connectors
[20] https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview
[21] https://claude.com/resources/use-cases
[22] https://claude.com/resources/use-cases/clean-up-promotional-emails
[23] https://claude.com/resources/use-cases/organize-files-in-google-drive
[24] https://claude.com/resources/use-cases/thoughtful-gift-giving-with-claude
[25] https://whitepapers.theregister.com/
Re: Boom
Nah. You pay the 100 bucks a month for a moronic employee to play with high explosives without any sort of safety gear or protocols and if he happens to blow himself up, it's all his fault.
You don't have to worry about firing him. Just sue the a$$ off the outfit that supplied the AI which was sold on the predicate that it'd allow people with minimal application skills to delegate activities beyond their competency to perform or check someone else's (thing's) work to an AI model with a different set of limitations.
Oh dear.
".....that allows people with minimal application skills to delegate activities beyond their competency to an AI model with a different set of limitations."
I'll finish that for them: "Unfortunately, since the task so delegated is beyond their abilities, they'll have fuck all chance of spotting something going badly wrong."
TL:DR: you're screwed.
And don't we just love having a helpful colleague always looking over our shoulder?
After leaving countless jobs due to being micromanaged, now AI wants to do it to me as well? Fuck that for a game of soldiers.
It's the count that thoughts
"thoughtful gift giving" by getting a machine to do the thoughtful part? How does that work?
All well and good
But wouldn't it be far more practical to just write (and run) a shell script that fires random commands at arbitrary times on one's computer (or on your boss/neighbor/little-brother's)? It would provide much the same entertainment surface factor at a way lesser cost in energy and water imho ... with a better profile in upholding responsible resource stewardship.
And it'd be a right sure fire hit with them self-driving (bumper) cars too ... like a delicious wrestlemania-inspired top-rope Markov-chain reaction pile-up wheelbarrow pyramid ... I mean, true genius-orama for your candy gluts, if you can smell it! ;)
As far as I am concerned, Anthropic Claude and all of his delusional, hallucinating AI mates can basically F&CK RIGHT OFF!
Then again, they do sometimes give us [1]a bit of a laugh when those who should know better make use of them (Copilot in this case) and embarrass themselves by treating the output as gospel!
[1] https://www.bbc.co.uk/news/live/c394zlr8e12t
Cow·orc ?
All the subtlety and agillity of a Jersey cow in an open plan office crossed with the congenital hatred and spite of Tolkienesque Orc ?
Hell ! I have shared office space with these cow·ork monsters long before agentic AI was even a twinkle…
Re: Cow·orc ?
The matching Nano Banana [1]logo sure is the cutest!
[1] https://simonwillison.net/2026/Jan/12/claude-cowork/#bonus-and-a-silly-logo
Oh lovely, another AI use-case aimed at deskilling jobs so they can (in theory) get just about anyone in to do the job and then pay them peanuts.
This has always been the case. Look at technical support roles? Used to be people who actually knew the product, how it worked, all the oddball quirks. Now; outsource it to a call centre in some part of the world and give them nothing more than a script to blindly follow.
This is nothing more than a continuation of that - short-term saving, more shareholder value, better ‘on-paper’ results, bonuses all round for the C-suite.
Well, like I said, short-term!
Boom
So you pay at least 100 bucks a month to play with high explosives without any sort of safety gear or protocols and if you happen to blow yourself up, it's all your fault.
What a great deal.