I don't think Firefox is immune to this, just that because its architecture is different, Anthropic didn't bother coding a bridge for it (given its market share).

The main issue here is that Anthropic violated one of the most important (implicit) tenets of applications in a computer: don't touch other people's shit. Claude.app modified Brave (and others) configuration, adding an extension without user consent. An extension that, by the way, gives full control of the browser to Claude, including reading the DOM for browser tabs unrelated to Claude (for example, the one where you just entered your credit card details).

This article is about Claude Desktop, the Electron-based macOS application

So you are bitching a chrome browser reconfiged your other chrome browser.....

Once again, brave browser sucks. Use Firefox or a Firefox fork if you want any real security.

Pretty solid take down. It could have been cut in half or a third, but all good criticisms.

It's hacking stuff but at what cost? Can I use it to direct an agent at my self-hosted stuff and see how secure (or insecure it is)? Or will that set me back thousands and thus make it only feasible/worth it for the rich and wealthy?

This is an interesting perspective. I have not thought about this even as a developer.

I used to be a developer, and I completely agree.

I don't owe anyone anything. And if you won't compensate me for work you demand, the less I am willing to cover your mistake.

“Supply-chain” is an invented capitalist digressive term that they forwent compensation for security. Even in our /c/, folks think capitalists will pay 7 additional days to review issues at no cost. It’s preposterous. Nazis prefer automating our quality assurance.

No pay, no game.

Edit! This type of ignorance even extends into other industries! Here’s my scene, making a bounty, not accounting modern Nazi costs of hardware.

Not updating with audit would work if every direct and transient dependency provided security updates for every version. But they don't. Often, security updates are for the most recent version or versions, and if you're far behind, you now have to audit a lot more.

Transient dependencies are an audit problem, too. To audit something, you have to essentially audit recursively. Many libs use many other libs of varied authors.

Our systems are too open, too vulnerable. A build or check being able to access all resources is a fundamental systematic vulnerability.

There is some truth to this - staying on the bleeding edge exposes you to things earlier.

Not really surprising, but maybe a consequence people who want to be on the latest version immediately did not consider. Good article!

But then your unsecure, cheap, and fast competitor will upload the features and solutions of your scope 14 days ahead of you, over and over again, until you're suddenly years behind the rest. Are this unaware of the modern business deployment model?

Do you expect there to be no crime and no money if capitalism is replaced? And effort toward mitigating supply chain hacks is definitely compatible with capitalism especially with such an easy mitigation as proposed in this article.

Nope.

Did you comment this on the wrong post?

incompatible with capitalism

Bullshit

Yeah, I'd like some third-party confirmation for all of Anthropic's claims about their latest model, thanks.

How about don't let ai read your anything.

I'd go further: have a separate user-account, on your own machine, whose entire filesystem is separate from the rest of your machine, for using LLMs in.

NO access to your normal accounts, your email, your browser-history or bookmarks, NOTHING except that-account's stuff.

Machiavellianism is intrinsic to the companies which produce them, we need to be presuming it to be intrinsic to the LLM's, too, as for some of them it is intrinsic, & we don't know which ones, yet.

Zero-Trust.

_ /\ _

Just sandbox it instead.

Instead, upload .env poisons.

Yeh, but it's serverless! Where else do the private API keys go, other than in the web page?
\s

The researchers found 1,748 active, verified credentials from major service providers (including Amazon Web Services, Stripe and OpenAI). These credentials were found publicly accessible within the live code of websites.

I mean, what else is new?

They are all going to end in a undisclosed amount paid with no admissions of wrongdoing.

Rubber stamping SOC 2, HIPAA, GDPR and ISO27001 for all but their highest profile clients. Lawsuits will be something to see

no cap