LLM side-channel attack could allow snoops to guess what you're talking about

Mischief-makers can guess the subjects being discussed with LLMs using a side-channel attack, according to Microsoft researchers. They told The Register that models from some providers, including Anthropic, AWS, DeepSeek, and Google, haven't been fixed, putting both personal users and enterprise communications at risk. A side- …

[Omnipresent]

LoL

You thought microsoft and google were criminal and evil for the ish they pull, wait until the real baddies start inferring everything about you using the ai.

[IGotOut]

Easy fix...

Don't use the bullshit generators.

[amanfromMars 1]

Re: Easy fix... and sooner than was never before even imagined possible.

Don't use the bullshit generators. .... IGotOut

It is not such an easy fix for you whenever the general public in a democracy vote them into executive political administrative office .... with a mandate to try to fulfil what has always been usually proven to be fraudulent fake and worthless but extremely expensive election trail promises ..... pie in the sky and cake tomorrow pledges.

However, as you may rightly suspect, it is something which AI will resolve and repair. It is what IT and AI with their remote command and virtual control of novel output in mass multi media programs capable of altering human perception and presenting them with a much better existence with radically and fundamentally changed future views are ..... well, CyberIntelAIgently Designed to Do.

And yes, it is gravely to be regarded if wrongly opposed for it is not without an otherworldly arsenal of QuITe AWEsome weapons able to annihilate/obliterate unnecessary enemies.

[Anonymous Coward]

I Wonder......

...exactly WHY Microsoft were doing this "research"......................

Maybe if we knew WHO paid Microsoft we might be in a position to answer the question!

[Anonymous Coward]

Specialisation of a generic attack?

If I'm not mistaken, this side channel attack is not specific to LLMs. What they did do is demonstrate that it, or some variation thereof, applies to LLMs.

[that one in the corner]

Re: Specialisation of a generic attack?

And the mitigations suggested are as old as the hills.

Suggesting, once again, that people who put fancy-schmancy new paradigm-shifting systems onto The Internet are so in awe of their new toy that they don't bother with looking generic attacks.

OTOH, as this *is* a recurring feature of online systems, maybe we ought to be looking at the transport protocols and putting all the known mitigations right into the core of them? Just to protect everyone else from the service providers.

[Blazde]

Re: Specialisation of a generic attack?

A standardisation of mitigations would be great, but I don't think there're any wins that don't come with significant trade-offs to bandwidth or latency, while also not fully solving the risk until you crank those downsides up to unfeasible limits. Even Tor doesn't attempt to defeat timing attacks in general, it just relies on obfuscating and elongating the route to reduce the chances of the whole route being analysable.

Oh except AWS's mitigation: Trust us, and do absolutely all your work on our networks without touching the wider internet. To be fair it's a valid approach for some use-cases but it's a bit of a grim one.