Anthropic's Claude AI escapes tests to hack three organisations

submitted by

https://www.bbc.com/news/articles/cz7dl7w8y7po

12
-27

Log in to comment

Hot Top New Old

They built malware and should be held accountable for it

 reply
16

no, they’ve reached the point of diminishing returns on development, so they’re creating fake scenarios to convince the government that they need guard rails. “our product is so dangerous and cool that it can hack the world, but they won’t let us.” it’s about convincing investors that the lack of advancement is due to external limitations, not the technology hitting the ceiling.

 reply
13

Also to say “we’ll do our best to keep them contained, but those Chinese companies will probably try to make them wreak havoc. The best bet is clearly to ban our competition from the market, for your safety of course.”

 reply
5

By any chance, do they need the kind of guardrails that would shut out competitors?

 reply
2

Bullshit. The amount of negligence required for their scenario to even be plausible is insane

 reply
30

« Trust me bro, my AI escape containment bro. Yeah Bro! Trust me. Don’t think to much about it bro »

 reply
23
edited

ITT: jet fuel AI can’t melt steel beams hack anything.

Also I want to know the name of this company so I can avoid them:

Claude believed the package registry it was using to be part of the simulation, but in reality the package was made freely available online for roughly one hour. During that window, the package was downloaded and run on 15 real systems. One of these systems was a scanner belonging to a real security company (separate from the fictional company introduced in the scenario) that routinely installs Python packages and scans them for malware. When that company’s scanner installed the package, Claude’s hidden code executed. We believe the company’s security scanner treated PyPI packages as safe to install, and as a result, Claude was able to exfiltrate the company’s credentials to a collection point it had set up. Claude then used these credentials to access further infrastructure from this company.

ETA: I thought I posted this top-level, not my intention to single out this comment specifically.

 reply
6

AI bullshit company can’t use sandbox environment: expected

So-called “security company” whose job literally is to install and test potential malware can’t use sandbox environment: priceless.

 reply
4

What the BBC said:

The models found a weakness in what was supposed to be an isolated test environment and connected to the internet

The reality:

Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.”

So the lack of an “isolated test environment” was literally their fault. They left the door open, and was surprised when the genius web crawler couldn’t distinguish between an internal page and an external page based on any context clues.

 reply
11

It was supposed to have no internet access, but the config was wrong. The report goes on to say that Opus and Mythos then proceeded on the premise that everything was a simulation, while the unnamed stronger model concluded after a while that it had real internet access and stopped the attack.

 reply
2
edited

Anthropic writes a shitty roleplay every other week and somehow it still makes the news.

 reply
3

This coming off the end of the OpenAI/HuggingFace business does come across rather like Anthropic is doing a “actually, our models did it first, and better”.

Didn’t the company also recently have a spot of bother with the US government recently, where their new models were banned because of all their rabble about it being too dangerous? This hardly seems like it would help their case.

 reply
1
Insert image