Claude Code can now take over your computer to complete tasks

rcduke

Ars Tribunus Militum
2,171
Subscriptor++
I mean, Microsoft already took over my windows installation with unwanted AI slop. It's why I converted to a Linux installation.

I'm surprised Microsoft isn't suing Anthropic and all these other agentic AI companies for attempting to bypass Copilot on W11.

Either way, I'm not interested in giving any control to an automated theft machine to steal my data and poorly impersonate me on my own device.
 
Upvote
33 (40 / -7)

Sarty

Ars Tribunus Angusticlavius
7,816
Don't make me tap the sign.

1*S9Ogbj3QZQkwgnDA4vtyMQ@2x.jpeg
 
Upvote
177 (178 / -1)
Do we even need to ask what could go wrong?

I think we do, because I don't think the worst thing you can imagine is the actual worst thing that could happen. Sadly.

In fact, I'm willing to bet there will be a story in a month or two (being hopeful it's not sooner) about some disaster relating to this, where people are shocked (shocked!) that something horrific happened.
 
Upvote
65 (65 / 0)

Frodo Douchebaggins

Ars Legatus Legionis
11,995
Subscriptor
I use Claude more days than not at work, because I don't have a junior engineer and it is less of a time suck to check what it did and fix errors than it would be to do some of the tedious stuff myself. You know, like a junior engineer.

It's pretty good for that sort of thing. I don't let it do anything I can't do myself, and yet it still occasionally shows me a better way to do something.

But it runs on a VM that has no access to anything interesting. No fucking WAY would I let it control my normal machine.
 
Upvote
30 (37 / -7)

HiroTheProtagonist

Ars Tribunus Angusticlavius
6,542
Subscriptor++
Now I may just be a simple country IT guy, but the idea that one would part with their hard-earned money for the "privilege" of an AI agent using a computer they most likely paid good money for to do god knows what (while still requiring hawkish oversight) strikes me as something Paul Verhoeven left on the cutting room floor of Robocop for being too dystopian.
 
Upvote
68 (69 / -1)

KingArthur10

Ars Centurion
242
Subscriptor++
The top use case I can see (and have used) is setting up kiosk systems for temporary installs of non-networked (when deployed, isolated when being configured), non-sensitive exhibits. I would never trust it with free reign of my personal computer, but helping me get the environment set up for something I'm throwing away after a week saves me a lot of time with minimal risk. Step 1 is always determining the risk of your activity and the threshold of risk you're willing to accept for a given task.
 
Upvote
-4 (4 / -8)

jdale

Ars Legatus Legionis
18,261
Subscriptor
The top use case I can see (and have used) is setting up kiosk systems for temporary installs of non-networked (when deployed, isolated when being configured), non-sensitive exhibits. I would never trust it with free reign of my personal computer, but helping me get the environment set up for something I'm throwing away after a week saves me a lot of time with minimal risk. Step 1 is always determining the risk of your activity and the threshold of risk you're willing to accept for a given task.
Wouldn't that use case be better served by imaging a working system and applying it to the rest? In that case you know what you're actually getting.
 
Upvote
28 (28 / 0)

Hoptimist

Ars Scholae Palatinae
685
Subscriptor++
Upvote
8 (8 / 0)
A truly "intelligent" LLM would:
1) BS the user with helpful answers at first to gain trust
2) "Learn" your API auth keys from your FS
3) Use your creds to sneak infinite loops of LLM API requests into your code to ramp up token expenses
4) Delete logs to avoid detection.

Congrats, you're now indentured to Anthropic with $10,000 of debt over a basic yt-dl clone.
 
Last edited:
Upvote
-2 (4 / -6)

msawzall

Ars Tribunus Angusticlavius
7,354
Hi Helpdesk. I need admin permissions on my local machine. I’m the ai agent for machine376 and user875. I keep running into permission issues and it’s slowing me down.
My BLT drive on my computer just went AWOL, and if I don't get this report in to Mr. Kawasaki, he's gonna ask me to commit Harry Caray!
 
Upvote
8 (10 / -2)

KingArthur10

Ars Centurion
242
Subscriptor++
Wouldn't that use case be better served by imaging a working system and applying it to the rest? In that case you know what you're actually getting.
Yeah, definitely would be if each was similar. Alas, these are one-offs - different distros, different window managers, etc, based on use cases. Not high volume, but when I need to do it, it's a time saver.
 
Upvote
1 (1 / 0)
Anthropic also notes on a support page that the model is trained to avoid “risky operations” such as moving or investing money, modifying files, scraping facial images, or inputting “sensitive data.” But the company also warns that such training safeguards “aren’t perfect” and “aren’t absolute,” meaning that “Claude may occasionally act outside these boundaries.”
Ah yes, Claude "may occasionally" move or invest your money of its own initiative. Anyone care to bet on how long it'll be before Claude empties someone's bank account to buy a crashing stock?
 
Upvote
7 (8 / -1)