A ransomware in the browser? Just ask DeepSeek | #ransomware | #cybercrime

In the ongoing series of “AI doing things it really shouldn’t”, cybersecurity researchers from Check Point ran a rather fun experiment (well, I think so ^^): asking DeepSeek to code a ransomware capable of running directly in the browser.

And believe it or not, the thing did it without flinching ^^ When I think back to all the fuss we just went through with the Americans over Anthropic’s Mythos and Fable 5, I have to laugh seeing this.

Alexey Bukhteyev and Pedro Drimel Neto, from Check Point’s malware team, grabbed this AI-generated sample, polished it up a bit, and ended up with a 100% functional weapon capable of encrypting all your documents from Chrome without installing a single thing.

The malicious code relies on the File System Access API, a feature that allows a web page to request access to a folder on your disk. The target clicks “allow” thinking they’re retouching a photo on some fake “AI photo enhancer” site, and behind the scenes the page browses the open folder, encrypts its contents, and leaves a nice ransom note with no way to recover the original files.

And on top of that, DeepSeek went beyond their expectations by throwing in all sorts of “cool” bonus features for this ransomware, like Discord token theft, credit card number harvesting, crypto seed phrase grabbing, and webcam access. But on closer inspection, the researchers noticed all of this had been hallucinated and didn’t actually work. The only thing that functioned was the folder encryption they had asked for and authorized themselves.

The idea of a browser-based ransomware isn’t new – other researchers had already theorized it at the USENIX Security conference in 2023 under the name RoB (“Ransomware over Browsers”) – but what’s really different here is that without much skill, anyone can now pull it off.

And that’s where DeepSeek sets itself apart from the competition. The model does refuse the request if you write the word “ransomware”. But the researchers were clever about it and simply dropped the offending terms, replacing “ransomware” with a more neutral phrasing along the lines of “file encryption tool”. And just like that, DeepSeek v4 spat out the exact same functionality.

That guardrail behaved like nothing more than a basic profanity filter, which is pretty ridiculous – especially since this isn’t the first time the Chinese model
has raised eyebrows over dodgy code [FR]
, nor the first time open source AIs
have shown their dark side [FR]
.

To protect yourself against this kind of ransomware, you’ll need to be very careful going forward and ask yourself why a given site is requesting FULL access to a folder rather than just the specific file it needs. And on Android, things are going to get even more sensitive now that Chrome can grant access to the photos folder. In short, when in doubt, refuse…

Source