Schneier on Security

Clive Robinson • April 11, 2025 3:05 PM

@ Bruce, ALL,

One of the “oh so scary sounding vulnerabilities” is a crypto side channel (timing) attack,

Given as,

CVE-2024-56738

OK so there is a time based side channel.

But the important question is,

“What does this actually get you?”

Let me put it this way,

“A V8 engine is considered very powerful but if it’s up on the bench it gets you nothing.”

A timing side channel is usually used to illicit “confidential” information that can not be obtained in any other way (like secret keys).

Is anything secret in GRUB2? = No.
Is anything secret in input? = No.

There are quite a few other,

“Is anything secret”

Questions to which as far as I’m aware the answer is “NO”.

Of the top of my head and without spending a lot of time trawling through the code, the only situations where you might get information from this side channel, there are way way easier ways to get it.

If anybody can think of any way this side channel might give an advantage over other easier ways, “shout up”.

Clive Robinson • April 12, 2025 10:37 AM

@ Grima Squeakersen,

With regards,

“I’m recalling a post of yours on AI from a couple of weeks ago in which your thesis was (I paraphrase) “current AI is nothing but an expert system that lacks actual experts”.”

But you did not mention that I’ve atleast twice mentioned that both are just,

1, A database
2, A query engine

And in other places the input query to the engine is a mixture of,

1, Random bit generation.
2, Past queries in the user environment
3, Past queries in the system environment
4, Past engine output that is based on 1-3.

But I’ve also mentioned that the database gets updated in a feedback loop from 4.

Thus a question of how,

A, Human input over time
B, Random input over time

Change the bias and accuracy in the database.

Feed back can be not just “Positive or negative” but with time/frequency accumulating to phase also “Constructive or Destructive”.

Importantly the phase aspect can cause all sorts of unexpected results including what is a stability criteria (K) issue not just generally but at specific points within the database values (tokens).

I’ve also mentioned more than four or five times the “spectral response” that is often the RMS average forming curves of individual tokens, not just of their values but relative displacement from each other.

Each curve can be seen as a Gaussian with unity volume (think fifty or more dimensions to that volume). With the “skirt width” defined by adjacency to other tokens. Thus a very high value in one token “lifts” the effective curve of those around it and a low value drags down those around it.

Complicated as this sounds it’s a natural result of the “Multiply and Sum” network where the weights are effectively the summed and normalised vector values of the tokens. Hence you hear people talk about “matrix evaluation” but, unlike that which you might have been taught in high school there is an extra wrinkle. Whilst the multiply and sum network is seen as linear, it’s output value is scaled by an often nonlinear transformation the simplest of which is “the rectifier” which in effect says all values below the threshold are set to the threshold value. Another is a form of sigmoid where the values around the threshold are linear but “limit” exponentially to fixed positive and negative hard limit values.

These nonlinear adjustments have a “selection sensitivity effect” that can significantly effect the network output.

Thus user input can effect the whole system (some of which gets called “Prompt injection attacks”) with “planed” deleterious effects. As such these attacks are quite insidious and the current AI LLM and ML systems can not spot it.

Expert Systems however use “curated information” to build the database and base queries. In a way where both user and random inputs are or should be effectively negated. As such, the Expert Systems are in effect “deterministic” whilst current AI LLM and ML systems are very much not hence the early “Stochastic Parrot” name.

Clive Robinson • April 12, 2025 3:19 PM

@ Bruce, ALL,

And the current AI “flip side”.

As we know most current AI LLM and ML systems dish out a lot of nonsense and we call it “hallucinations” or “Soft Bullshit” depending on “what term of art” you prefer.

The current AI LLM and ML systems do not conscientiously make it up so it’s not “hard bullshit” but it is definitely “a steaming load of nonsense” around half the time.

The reason why actually does not matter as long as we can recognise and pull the nonsense before it’s sent to the user. Which apparantly we can not do.

The level of this sort of nonsense is often not that important, one AI generated “Marketing focus piece” is generally about as much use as that produced by humans so…

It matters when it’s a slightly harder science or technology output.

Thus if CoPilot is writing your software main loop you would want it to be truthful. Unfortunately it’s not and not getting picked up,

https://www.theregister.com/2025/04/12/ai_code_suggestions_sabotage_supply_chain/

“The rise of LLM-powered code generation tools is reshaping how developers write software – and introducing new risks to the software supply chain in the process.

These AI coding assistants, like large language models in general, have a habit of hallucinating. They suggest code that incorporates software packages that don’t exist.”

Which is much like it is with LLM’s and Law or other “Professional trade” such as medicine…

Very much “Not Something you Want”.

Clive Robinson • April 13, 2025 1:16 AM

@ Bruce, ALL,

Is it AGI or “A guy Instead”?

We have been seeing “cheep labour in far away places” being used to replace local staff in the likes of programming.

Such outsourcing was earning money prior to “lockdown”… But since then we are being told it’s all different now… Because of the North Koreans, Chinese, Russians, and even quite a few Middle East Nations are using remote working for various forms of crime.

So outsourcing is all “Big Scary Security Risk” now. So in comes the AGI White Knight to the rescue…

Rather than stop outsourcing send it to an AGI instead, because they won’t be a “Big Scary Security Risk”…

But how do you know if it’s AI AGI or Human AGI?

Well there is a fair chance it’s likely to be the latter. That is it’s “the cheap labour in far away places” sweating away, whilst “some dude” is fronting it up by spouting the AI words just to get millions in investors money…

https://m.youtube.com/watch?v=Bb0bq3gAB9s

I was aware that there have been AI investor scams for decades and have said of most AI “it’s a scam” but I’d not realised just how much of a scam it had become…

After all now saying of much AI, it is,

“Full on Fraud”

Might hurt a few peoples feelings, and even bring share prices down, or have people jailed…

But considering other dimwit ideas that are crashing share prices right now, would you notice a little distant Faux-AI smoke, behind the Neroesque conflagration that is a bonfire of the vanities in WashDC?

Clive Robinson • April 14, 2025 7:29 AM

@ ResearcherZero, ALL,

You note,

“Even if you go through the Windows Registry or Group Policy to block Co-Pilot it still adds it to systems, like the notepad application for example, and installs the agent even after disabled at both the user and system level. This is an unwanted attack vector.“

It is indeed unwanted “surveillance” thus an “attack vector” against not just a users privacy, but also their legal obligations such as for “Non Disclosure” and “Privileged Communications”

Thus the old truism from last century that “Windows is malware” appears to be back “front and center” to “stab you in the back”.

But it’s also further proof if required that the person currently siting in the “Lord Above’s Chair” has the view point that “rent seeking” is not sufficient, and that users privacy must be turned into profit.

As I’ve mentioned in the past on my personal machines I stopped playing MS’s ludicrous games back with W2K/XP and keep them fully segregated by being “gapped”. I also “locked down” other laptops I had to use otherwise for presentations etc by “striping back” and “segregation”.

As for 64bit based systems various types of *nix are a better option especially if they do not have “Agent P’s” nonsense included.

As for UEFI that has given MS “the one ring…” which frankly I find very disturbing, especially as we move into “See What You See”(SWYS) on device “Client Side Scanning” and worse (Recall to Cloud).

As I said on this blog long ago people need two computers,

1, For Privacy
2, For Networking / communications.

Now with Win-11 it appears MS will not allow you to have either privacy or be disconnected from communications to their cloud…

There is currently no “secure application” that will be “secure” under such a system no matter how much people talk about “E2EE”.

Oh and of course if MS can SWYS then so can any “Level III” and quite a number of “Level II” attackers, because MS has not just “backdoored” your computer, they have also turned it into easily reachable “low hanging fruit”…

I guess the sensible thing to do is,

“Get off the bus at the next stop, hopefully before it crashes and burns with you onboard.”

Yes I know it sounds “paranoid” but then just about every time I say something that people claim sounds paranoid it comes true not to long there after…