Schneier on Security

Clive Robinson • August 15, 2026 6:10 AM

@ r, ALL,

Like many others Matthew Green has the history and facts of what eventually became a manufactured nonsense by the FBI wrong, if not totally backward…

I tried to “make a post” on his site using the “email option” but it just disappeared…

But the real reasons are,

Firstly the US got politically embarrassed by the Soviets and thus made the stupid mistake way back last century of cutting back with “boots on the ground” HumInt “methods and sources” and instead went down the technological rabbit hole of ElInt etc and “watching from space” (which is now rather more in danger of “going dark” due to anti-sat missiles etc).

HumInt when conducted properly is orders of magnitude better than ElInt etc. So the US became overly dependent on it’s diplomates tittle-tattle[1] and Sovereign Nation partners HumInt sources and methods (the UK being but one[2]).

Secondly Law Enforcment got fat and lazy again back last century long before the “public internet” became a thing in the later 1990s. And it was the 5th FBI Director Louis Freeh that started the nonsense that much later became “Going Dark”.

As I point out from time to time the work of Claude Shannon, and Gus Simmons gave proof of both “Perfect Secrecy” and that it would also be “Covert” way before that. And I’ve described here in the past in detail exactly how to do this with just a pencil and paper OTP and phrase Code Book, basic technology available from before WWII and in everyday use just about everywhere in WWI, so well understood in use.

The real issue is I also point out from time to time is,

“user convenience”

and Law Enforcement reliance on people taking the “easy route”. That is ordinary people going for “convenience” rather than the minor inconvenience of OpSec that would give them guaranteed Communications Security.

In effect the FBI and other Guard Labour have become,

“Donut munching keyboard warriors with ever expanding waistlines and need of type II meds, scared whit less of what the public might do to them if they chose to.”

The thing is there is plenty of proof of two things with criminals,

1, They want hi-grade privacy.
2, The less smart ones want convenience more.

Just read up on the fake secure mobile phones that so many got caught by like “Encro-Chat”,

‘https://en.wikipedia.org/wiki/EncroChat

And how the whole operation got blown due to a “gym-bunnie”…

‘https://www.essexnewsandinvestigations.com/single-post/corrupt-police-analyst-who-leaked-encrochat-op-was-promoted-after-moving-into-suspected-crook-s-home

Yup you could not make it up… But this is the real thing that scares not just the Guard Labour, but their political bosses that want everything on the cheap or for free.

[1] What Chelsea Manning was accused of giving to Wikileaks via it’s founder and public face Julian Assange.

[2] Because of US politicians “burning” other Sovereign Nations HumInt “Methods and sources” for “press sound bites” etc (most notably the UK and Dutch). The US is finding that they are nolonger considered “trustworthy” and this means “no raw intelligence”. Which is a significant issue as anyone who remembers the “Yellow-Cake” nonsense will know,

https://en.wikipedia.org/wiki/Niger_uranium_forgeries

Clive Robinson • August 16, 2026 8:50 AM

@ Bruce, ALL,

Do they never learn from DRM fails?

Some of us remember the “Digital Rights Management”(DRM) “Watermark wars” from the end of the last century and how miserably they all failed. From what I remember what finally put the nail in it was Prof Ross J. Anderson and his UK Cambridge Computer labs group.

They used a simple trick to 2D twist the image, sufficient to destroy the DRM but low enough not to be too perceptible to human eyes.

So imagine what a wry laugh I had reading this,

Anthropic says text watermarking scheme relies on inconsequential words

‘Shall I compare thee to a summer’s afternoon’ is the sort of thing this will make, and others look likely to adopt it

“In an effort to “watermark” text that Claude has generated and comply with the EU AI Act, Anthropic unveiled a plan on Friday to modify its bots’ choice of words in a way that would be detectable as the product of an AI.

Traditional watermarks are patterns or images overlaid on currency, postage, or official documents as an assertion of authenticity. In the digital realm, the term is more flexible and can refer to a variety of techniques for applying an identifier to electronic data.

Anthropic’s approach involves influencing inconsequential word choices made by its models, a technique introduced in Google DeepMind’s SynthID-Text paper.

To oversimplify things, large language models are fancy autocomplete engines which work by predicting the next word in a sequence of words. Anthropic explains that while composing sentence output like “The weather today was cold and…” a model like Claude might respond with words like “cold” or “gray” and would be unlikely to respond with a word like “sugary.”“

And entirely miss the point of why it will fail miserably…

To be used as a watermark it has to have the equivalent of a “stream cipher key generator” used as a “feint stream cipher”.

That is a “stochastic source” is needed to bob back and forth over the peak probable word curve. It’s what LLMs do any way hence “stochastic parrots”. The thing is it has to be a very special type of “stochastic” to work as a Watermark.

Without going into the details the watermark becomes a “stream cipher” and thus inherits the failings of the “One Time Pad”(OTP)…

Some of the thing people are told about Stream Ciphers are,

1, Never use the same Key-Text, with different Plain-Text
2, Never send the same Plain-Text with a different Key-Text.

As this opens them up to sometimes quite trivial attacks.

It is however,

3, OK to send the same Plain-Text under the same Key-text repeatedly.

Even though it is unwise to “re-encipher” rather than just copy the cipher-text as simple errors can arise (which is how the WWII German high security “fish” cipher was completely broken, and modern computers came about as a consequence).

Thus a stamp or other static image or text can be watermarked / encrypted safely the once, and THEN copied as many times as you like quite safely.

But IF the image or text changes and you use the same or similar Key-text then you can not re-watermark / re-encrypt it with the same or similar key text without leaving it open to easy attack which would include “easy forgery” or “easy removal” of watermarks. Which makes the entirely worthless.

But… Some will argue that as long as the Key-text is

“sufficiently different”

It will all be OK…

The actual answer is “probably not” and turns on the fact that to function as an “independently verifiable Watermark” the Key-Text can not be “random”…

It requires as a minimum an inbuilt near-unforgeable distinguisher used as a “sentinel signal”. Such things are these days based around “One Way Functions” for good reason. But they also drag a whole world of problems and potential insecurities with them.

So any “simple system” Anthropic et al are likely to come up with will have a high probability of being “not fit for purpose” in fairly short order…

Something that really should not surprise anyone paying attention.

And something the rest of the article takes “The micky out of” in other areas…

I’ll be blunt and say in a watermarked way,

“Failure is now an outcome”