Schneier on Security

Clive Robinson • April 11, 2025 6:16 PM

AI no closer to the magic G spot…

Stanford University’s Institute for Human-Centered AI (HAI) has produced a weighty report on current AI LLM and ML systems that basically says for all that extra coal in the furnace and water in the boilers it’s realy not getting anywhere near close to the hyped up G spot.

Where there has been improvement it’s in benchmarks, and as we all should know by now… The LLM gets right what it has in it’s corpus… So in a way the LLM improvement is like a cheating student using somebody elses A+ paper.

To save people the strain of getting to grips with the report it’s self, The Register has what reads like a human generated synopsis,

https://www.theregister.com/2025/04/11/stanford_ai_report/

The important thing to note –other than it is not making returns– is,

“[T]he report also stresses that complex reasoning is still out of reach for AI models. Even with mechanisms such as chain-of-thought reasoning to boost their performance, large language models (LLMs) are unable to reliably solve problems for which a solution can be found using logical reasoning, making them unsuitable still for many applications.”

Thus where they will show returns is by automating “make work” and similar jobs that actually serve not much purpose other than needlessly increasing cost. But… as “make work” is what the “Return to Office” senior management commands are really all about… Perhaps it’s managers that should be most alarmed about “AI replacing workforce”.

Clive Robinson • April 12, 2025 10:35 PM

@ Who?, ALL,

“The only reason I see to declare these technologies “legal” is that they are sponsored by the government.”

It rather depends on what you mean by “legal”, “sponsored”, and “government”.

Developing such systems based on statistics are not “unlawful” just about anywhere in the world where there are people with the technical capability.

This “nonsense” has almost always be developed initially as “anti-terrorist” etc and is in response to the likes of PIRA and other long term groups being sufficiently good at Fieldcraft and OpSec that traditional investigating techniques really don’t work.

Why do I say “nonsense” well it’s all based on the idea that,

“All such problems can be solved with more data…”

They can not for reasons a high school student should get taught, but for some reason gets forgotten when “Money comes a’knocking”.

Put simply the fact that a group of people all get classified by their actions as “undesirables” and are found to all have certain common traits… In no way means anyone with those traits is an “undesirable” or that all future “undesirables” will have these traits.

Such is the nature of “freewill” and “agency” breaking observed “cause and effect”.

There is even a quite old expression with respect to it,

“A wolf in sheep’s clothing.”

Basically as an “undesirable” with the capability of reasoned thought you look to “hide tells” that you are aware your opponents may be looking for, so as to render much of the quite limited investigative resources wasted.

Thus you get into the “secret sauce” mindset as an investigator, where you try to keep “the tells” you use secret. When you think about it given time and picking people up what ever your “tells” are they become known to the undesirables.

So what as an undesirable do you do?

The simplist thing is to not just make the tells act against the opponent but actually be incorrect or worthless.

There are two ways as an undesirable you to do this,

1, Not have the tells.
2, Make the tells to general.

It’s the premise behind the statistical game known incorrectly as “signal to noise ratio” taken from communications and information theory.

The primary assumptions behind “signal to noise ratio” is that

1, The noise has a known profile.
2, The source of the noise does not change.

There is a kind of hint to this by the use of the term “AWGN” that stands for,

“Additive white Gaussian noise”

https://en.m.wikipedia.org/wiki/Additive_white_Gaussian_noise

I could go down the rabbit hole of “statistical mechanics” but there is a simple experiment that nearly every high school student sees atleast once to demonstrate amongst other things “Brownian Motion”.

You start off with a quantity of water and divide it into two lots. To one you add a quantity of potassium permanganate that turns it a very deep purple hue.

You have a simple thermal still arrangement using two large round bottom flasks the bottom of which holds the now purple water and the top the still clear water.

You apply heat to the bottom flask and some warm purple water ascends into the top flask and some cool clear water descends into the bottom flask. It looks very pretty and is thus quite memorable.

The eventual result is a near equal mixing of the two samples of water.

The experiment is easy to make go in this forward direction… But ask yourself how you would get it to go in the opposite direction. That is how you get the well mixed in potassium permanganate to unmix into just the bottom flask but importantly remain well mixed there…

Practically it’s impossible, as it is in effect a physical “one way function”.

Now consider those molecules of potassium permanganate as being a model of “the tells” behaviour mixed in a general population…

There is a thought experiment called “Maxwell’s Demon Paradox” from a series of letters from physicist James Clerk Maxwell to Lord Kelvin back in 1867 to in effect violate the second law of thermodynamics,

https://en.m.wikipedia.org/wiki/Maxwell%27s_demon

“Lot’s of Demon’s have been proposed and so far they have all been deposed”.

The thing is “probability” says that you do not even need a Demon there is some very small but finite chance the molecules will spontaneously rearrange back to the starting conditions. But it probably won’t happen in your life time or the lifetime of the universe.

But the trick that the people rely on to sell such systems is,

“If you reduce the target population the probability comes up in any given period.”

Consider the case of just one molecule of potassium permanganate, it can only ever be in one of the two flasks. Therefore simple thermal circulation will pull it around from one to the other and back again much like those 1960’s “lava lamps”.

Thus if you rig the experiment with hidden heat sink/sources you can change the flask the molecule is in at which point like all good magicians and con artists you throw back the curtain and cry “Ahha”, collect your performance fee and scuttle off into the night before the audience works it out and forms a lynch mob or just pretends “they knew all along but did not want to spoil the fun for others”. Either way the money is over the hill and far far away.

Why do I know this… If you look back on this blog quite a few years now you will find me telling of how I inadvertantly got mixed up in such nonsense in the 1990’s in a University close to London, that were doing early “behaviour analysis” by CCTV to spot “undesirables” –beggars– loitering on London Underground Platforms.

It worked sort of because “beggars” can not really change the way they ask people for money so it stands out as a behaviour. But they could change where they asked for it… So the beggars moved from the platforms onto the trains, where they still are three decades later.

It’s just one of the reasons I’ve pointed out over the years here and in numerous other places that “CCTV for Crime Prevention” has a predictable curve of failure. First it works as it is intended because street crime “there” goes down as the criminals get caught and the smarter ones move out to where there is no CCTV. But then the crime goes back up again… the reason is CCTV is static in it’s capabilities, successful street criminals are quite dynamic in their behaviours and so effectively evolve around the CCTV in various ways so it becomes useless.

Eventually the authorities “caught on” to what I was saying, hence such systems are made not just “hidden” but highly mobile. Thus trying to be more dynamic than the criminals. Unfortunately there are ways that criminals can out evolve that. I won’t go into details of “flash attacks” and “steaming gangs” you can look them up or study medieval warfare techniques.

It’s one of the reasons London’s Met Police were accessing “mobile phone” data around the time of “The Croydon Riots” and later they took it up a notch or ten with EncroChat phones. As did the US FBI, and Australian, French, Swedish and many other Nations equivalents.

The thing is the Police unlike the Intelligence Services are limited in what they can do. To succeed they have to get successful “criminal convictions” which means two things,

1, The evidence has to be within Court Rules.
2, The “sources and methods” have to become “known” by lawful useage.

Which means smart criminals can stay ahead on the evolution curve…

It’s in part why I talk about,

1, Data : The message.
2, Meta-data : Message routing.
3, Meta-meta-data : User behaviour.

All are amenable to statistical pattern recognition be it cryptography, Traffic Analysis, Contact tracing, etc by the use of exponentially increasing levels of technology. As for the criminal “flattening the statistics” gets harder and harder as long as they remain within the technology constraints.

Eventually successful criminals will find ways to “use technology” beyond the constraints of consumer/commercial” technology use, where “billing” and “profit” by technology operators means user usage is auditable thus traceable.

One aspect of this was to use “burner phones” but whilst they stopped “Meta-data” analysis they did not stop “Meta-meta-data” analysis.

There are ways around Meta-meta-data analysis and I’ve mentioned one or two in the past. The problem with “communications” is,

“They are used by humans, and to err is human, and that provides the cracks, and it takes only one crack with collect it all.”

The thing is that technology does not have to operate in the consumer/commercial “technology sphere” where every thing is recorded… A look at the use of technology over at the Eastern edge of Europe tells you that there is much room for the more agile participants to stay well ahead. And yes I’ve mentioned some of this in the recent past.

It does not take a genius to realise that the agile party will always exist and when based on “smart behaviour” will out evolve the less agile party. Thus the only thing that will stop it is when the costs become to prohibitive for one side, giving a default win to the other.

Some authorities have realised this so have changed legislation such that a 3rd “commercial” party picks up their bill. It’s what CALEA and other legislation is all about.

But the price that is payed is that commercial interests are like “factory fishing” that very nealy made wales extinct and is now making all other edible fish stocks endangered.

Society is the sea humans swim in, and “commercial phishing” by corporates are destroying not just society as we know it, but mankind as well…

Clive Robinson • April 13, 2025 7:04 AM

@ Bruce, ALL,

Heads up : DragonFly BSD now has native DM-Crypt Disk Encryption.

[M]erged this past week was dm_target_crypt_ng, a next-generation implementation of their DM-crypt code for disk encryption.

DragonFlyBSD developer Michael Neumann re-engineered the DM-crypt code for this BSD operating system as this transparent disk encryption implementation compatible with Linux’s dm-crypt.

https://www.phoronix.com/news/DragonFlyBSD-DM-Crypt-NG

It now contains some of our host @Bruces crypto algorithms,

“Follow-up patches have refactored the crypto ciphers, improved the crypto cipher API, and introducing Twofish/Serpent CBC/XTS ciphers. Some nice improvements for those wanting to make use of disk encryption with DragonFlyBSD.”

Clive Robinson • April 14, 2025 6:09 AM

@ Bruce, ALL,

New method of no gradient AI ML

This is one to watch, but I’m not yet sure if you should do so with a bowl of popcorn or not (though it can’t hurt 😉

As some may remember, some years ago the “transformer” gave rise to the current AI hype bubble, that is now deflating for various reasons.

Not least is that it does not deliver in certain respects, nor is it ever likely to do so, outside of it’s current known capabilities.

I’ve previously described the DNN as being a DSP network that forms an “adaptive filter” and whilst some bristled at this, it’s now a more accepted view point.

The general use of DSP adaptive filters is to “reduce noise and distortion” thus increase the “perceived signal to noise ratio”. An associated term for this is “denoise” (it’s simple enough to understand but long winded to explain so I won’t do so for now).

Any way a new AI method and thus name as a term of art has popped up in a paper where the title gives it,

“NoProp: Training Neural Networks without Back-propagation or Forward-propagation”

The authors in at the start of the introduction describe the current ML method based on the transformer as,

“The canonical deep learning approach for learning requires computing a gradient term at each layer by back-propagating the error signal from the output towards each learnable parameter. Given the stacked structure of neural networks, where each layer builds on the representation of the layer below, this approach leads to hierarchical representations.”

As can be appreciated this method is not efficient and so consumes a lot of CPU cycles by the repetition involved with the backward and forward feedback to propagate changes to the network weights. In effect a laborious smoothing or low pass integration based process returning less and less with each cycle.

Not immediately obvious but nether the less true, is that such propagation can not exist in known biological systems. So the transformer or other back propagation methods are not representative of known biological learning systems.

They authors go on to say,

“[W]e introduce a new learning method named NoProp, which does not rely on either forward or backwards propagation. Instead, NoProp takes inspiration from diffusion and flow matching methods, where each layer independently learns to denoise a noisy target. We believe this work takes a first step towards introducing a new family of gradient-free learning methods, that does not learn hierarchical representations — at least not in the usual sense.
“

https://arxiv.org/abs/2503.24322

It needs be said this is not the first ML method that does not use back propagation, but it certainly appears more efficient.

It will be interesting to see what effect this has on current AI ML transformer method based systems and the companies that are so reliant on them…

Clive Robinson • April 15, 2025 8:25 PM

@ MrSmith, ALL,

What gets burnt with a burner?

With regards an article in The Register that says,

“The European Commission is giving staffers visiting the US on official business burner laptops and phones to avoid espionage attempts, according to the Financial Times.”

You say,

“Apparently I’m not the only one with this idea to use “disposable” phones.”

Using burner phones is not a good idea these days nor has it been for around two decades.

There is a special kind of “OpSec” you have to follow with them and most people really can not do it reliably even if they know their very own life might well depend on it.

Look at it this way, diplomats and their non-intelligence diplomatic staff are very easy to locate by the “Mark One Eyeball” or CCTV images during their normal activities. Likewise when abroad their social activities.

Cross corelating the eyeball sightings with traffic info freely available via “Signalling System Seven”(SS7) is very trivial.

As I found in the early to mid 2000’s when developing “traffic census systems” it’s actually very hard bordering on impossible to anonymize phones. Put simply they have two unique identifiers which you can view as a unique SIM ID and a unique Phone ID, along with Service Provider IDs, that you can trivially track in a whole number of ways including the cell site pass over protocols.

Thus the likes of the CIA/NSA will have a list of fairly valid identifiers of the burner phones within an hour or two of them being turned on. Especially if the Service Provider ID is “non-native” or one known to be “favoured” by a diplomatic station administration.

They are “100% Painted” the moment they contact any other service, phone or SMS provider. Or if the phone is supposedly Smart and uses the Data Network to do “chat” etc (think Apple and Android OS’s).

These days all Smart Phones have way to much easily identifiable Meta-Data and untrained or tired staff generate far to much easy Meta-Meta-Data.

Further as the CALEA interface gives Data in the form of audio and messaging “back haul” in “plaintext” form that will be easily obtained from “Collect it all” NSA or even FBI surveillance.

The diplomatic staff were actually safer when they could use “phone boxes and DTMF codes” and as they were still around back then broadcast based “pagers” that did not give away a persons location.

Yes there are ways you can use both dumb and smart mobile phones but you need much heightened OpSec.

As the CIA found out the hard way, trying to “get smart with the Internet” probably got something well over 20 Chinese Agents they had recruited executed or worse by Chinese counterintelligence services back in 2010 and things only went down hill from there for a couple of years till they changed handler methods sufficiently…

Clive Robinson • April 16, 2025 6:49 AM

@ Bruce, ALL,

This has become dummer than dum.

I mentioned not that long ago that those behind PKI Certs were likely to take things to ever shorter renewal periods…

Well even I was shocked to see this,

https://www.digicert.com/blog/tls-certificate-lifetimes-will-officially-reduce-to-47-days

It’s less than seven weeks and realistically when you add in likely snafu delays you will have to plan on the reality of just about a month… Even with “automation” it will be tight, without it will just go wrong due to winter maladies and summer breaks.

But “autonation” as a rule of thumb makes an attackers life easier…

So if anything, this change is going to create more security failings than it might potentially stop…

Clive Robinson • April 16, 2025 11:26 PM

@ ALL,

How many balls has Microsoft to drop?

Basically Microsoft’s various “Cloud Services” are failing repeatedly to deficient Microsoft work practices.

If this is in consequence of the Microsoft “Mega Surveillance” push of,

“AI Phone Home, with everything”

Or not is yet to be determined but the answer is likely in all probability to be in the affirmative.

So Microsoft appear to be reverting to “habits of old” of push out crap fast and maybe fix it before the next major release (or probably not).

The latest Yesterday –just one of several major SNAFUs recently– hit file sharing in Teams…

https://www.theregister.com/2025/04/16/microsoft_teams_file_sharing/

To some a “who cares” and they might not even have even seen the Microsoft ball drop.

To others however their production work flow came crashing down… Thus they were left scrabbling around trying to find emergency mitigations to the significant Microsoft “Ball Drop”. That as is the nature of such “Quick Hacks” would have effected not just the work flow, but the “security” of not just the emergency mitigation, but the down the road security of the organisational systems and data.

Apparently Microsoft are claiming they’ve “rolled it back”, but should we believe them?

That is, is the “roll back”,

“In whole or part?”

Something tells me it will at best be “in part” just enough to make things “look right” again at the surface… But with deep changes left hanging for “Certain more enterprising types” to “find and exploit” to their advantage not the Microsoft SaaS customers (think “Malware as a Service”)

Interestingly “The Register” article takes a similar,

“Don’t put all your eggs in the Microsoft basket!”

View point, listing some of the recent major Microsoft “Cloud SNAFUs” and then some “Microsoft PR guff/nonsense” befor as the last two paragraphs,

‘Matthew Hodgson, CEO of the [–Org name removed–], told The Register: “The Microsoft outage once again highlights the risks of relying heavily on centralised cloud services.”

He noted that going down the centralized route can create single points of failure. In this case, a back-end changed borked the files experience in Teams for many users. Hodgson suggested that companies could mitigate such risks by reducing reliance on any single provider through distributing data and workloads over multiple nodes.”

So the advice some of us were giving “long long ago” when “SaaS” was just a sparkle in a marketing droids eye. Back when we were saying “Service Level Agreements” were worth the bumf (toilet paper) they were printed on…

Clive Robinson • April 17, 2025 3:11 AM

@ Bruce, ALL,

This ARS Tech article superficially looks like it is “political”. But it’s actually not in the sense of “Trump Tariffs” that are just,

“Slapping a sticking plaster on for broken bones”

What it tells you is importantly why the bones got broken in the first place,

https://arstechnica.com/tech-policy/2025/04/14-reasons-why-trumps-tariffs-wont-bring-manufacturing-back/

I’ve worked in Engineering & Design one way or another for much of my working life, Manufacturing appears greatly in both disciplines.

That is from “one off” prototypes or art pieces through mass manufacturing of “Fast Moving Consumer Goods”(FMCG) where tens of thousands at a minimum of identical products have to be made in minimum time at acceptable minimal cost.

Surprisingly few in the West actually have the first clue about the process that produces everything that surrounds them every day of their lives.

Even if you live entirely “off grid” growing your own food whilst wearing homespun clothes, the tools you use that “force multiply” your efforts will have been manufactured[1]. By you or someone else, and the materials used sourced via a supply chain[2].

Even a basic nod to understanding the complexities involved in manufacturing is necessary for people to protect themselves from small or major events in their lives.

[1] The word “manufactured” originally ment “made by hand” by someone who was artisanal thus skilled and you “paid” them rather than do it yourself.

[2] Be if Cheese for food or casin plastic for buttons and knife handles, the start of the supply chain is difficult to see. In part because it’s so intertwined. Casin is a milk protein and you get it from mammals by husbandry and milking. Part of the husbandry is giving them food such as a bucket of grain and the likes of beets both of which are grown in enriched or fertilised ground. The soil enrichment obviously has it’s own supply chain, part of which is herbivore products such as waste and bones. So you in effect have no start to the supply chain.

Clive Robinson • April 17, 2025 3:52 PM

@ ALL Windows 10 users,

Resistance is futile you will be upgraded to 11

Due to a “pre-emptive” cods-up Microsoft have revealed that they can force a Win-11 upgrade on a user, despite all admin settings saying it should not be done!

We’ve seen Microsoft force unwanted upgrades in the past such as with the early Win-10 nonsense. Which not only upset a lot of people it also cost others –on “pay for mobile data plans” that were both common and expensive or were abroad– more than it would have cost to buy a then top of the line large screen laptop…

It would appear Microsoft is trying to talk it’s way out by saying,

“The problem, according to the mega-corp’s advisory on the issue, is that “a recent service change uncovered a latent code issue, causing impact.” Not all users are affected, but those who are will see an invitation to upgrade to Windows 11 regardless of how their IT admins have configured Intune, Microsoft’s own PC management tool.”

https://www.theregister.com/2025/04/16/windows_11_bypasses_admin_polices/

But is that honestly true?

I and others feel it’s very definitely not true, because it took Microsoft quite a few steps to enable what was “hidden functionality” built into their own management system…

“This latest SNAFU, which has yet another bit of dodgy code at its root, has taken Windows 11 as a feature update and bypassed Intune policies to prevent installation. Unlike the Windows Server 2025 fiasco, this appears to be going through Microsoft’s own management platform.”

Make of it what you will but I would not be surprised in the slightest if at or shortly after Win-10 going out of service you get hit with an enforced upgrade that may very well break your functioning system if not entirely “brick-it” beyond most peoples ability to repair (happened with the forced Win-10 upgrade).

Microsoft are sending out signals that you will not have any choice, you will have, forced on you,

1, Win-11
2, Microsoft “surveillance by AI”.

There are only a couple or three ways out…

The first is to stop using Microsoft entirely which all things considered would be the best option if you can.

The second if you can is to disconnect your computer from all external communications so it can not upgrade any part of it.

Thirdly is have a backup policy that lets you just overwrite the entire Microsoft OS and if necessary Apps. Likewise if necessary set the date back to “the last known good time”.

Fourthly if you have the ability go back before Win-10 or to the early versions of Win-10.

There is another option but potentially it’s not within Microsoft’s User Agreement…

Which is install say Linux with Wine, and then install Win XP, 2000, Win-8 compatible apps to run in it. This has the advantage of not needing an MS OS licence. However if you have an OS licence and CDs/DVDs then you can use a “virtual system”. Back a decade ago “Virtualbox” –from Oracle– was one such system, but best avoided where you can.

Last time I looked Wine had got just under 30,000 apps runing on it satisfactorily and more pop up on a regular basis, see,

https://appdb.winehq.org

And keep the link in your bookmarks 😉

If you are going to install Linux, a round up of various opinions is “Go with Mint (I’m an old school CLI user so have my own odd preferences).

However not all users are up to dealing directly with Wine, which due to it’s needed flexibility can appear a bit complicated. There is a way out of this 😉 When you buy real wine in a shop it usually comes in a glass bottle. Well theres a Linux tool / application which is a “visual wrapper” for Wine called bottles. This is the most recent article I could find that talks about it,

https://www.howtogeek.com/running-windows-apps-on-linux-with-bottles/

Where you can I’d advise against using virtual systems because commercial interests in that space are creating a lot of FUD and other nonsense at the moment. No doubt in a bid to extract large quantities of extra money to make their recent VC type investors happy as quickly as possible.

One way that uses a commercial backed VM offering is described in,

https://www.windowscentral.com/how-setup-windows-10-virtual-machine-linux

(Note the DRM work around given in the “Bypassing Windows 11 installation errors” section).

Thus avoiding a commercial VM is if you can do it probably preferable. I’ve used earlier versions of QEMU,

https://www.qemu.org

That I use as an “emulator” to run verisons of MS-DOS and Win3.x to run Borland or Small-C K&R compilers under. In either 32bit or 16bit x86 CPU emulation, to support some really really old software I developed last century that the “user” does not want to stop using because “his fathers factory needs…” (The things we do for friends and sometimes family 😉

Unlike many people I’ve got full “payed for” licence for those earlier MS OS’s and Apps not a “bought with pre-instaled” licence that most have. Importantly the printed licences that comes with them does not prevent running them in either a virtual system or under Wine.

Just to show what an old grumpy I am I still prefer Win XP or for some odd development software Win 2000.

I’ve been doing the equivalent of Wine since the mid 1990’s, it was with the Consensys version of AT&T Unix Sys5v4 and DosMerge. That as I’ve mentioned before enabled me to run multiple 8bit processor ICE development systems One a 486 with 8port serial card. The big problem was that Motorola had developed their software so that each ICE normally needed it’s own 286 or better MS-DOS machine, which when you need six at the same time… I had it setup “in the lab” that had “air con on chill” and talked to it via a network cable to do software development from the comfort of my pleasantly temperate office desk (I also ran test instruments that way as well).

The thing is people have been “brain washed” into “MS is the only way” nonsense and now Microsoft are clearly abusing it via “rent seeking SaaS Cloud” and “AI privacy destroying surveillance” –via CoPilot and Recall– neither of which people should be trapped into.