Hiding Capabilities in AI Models

Interesting research: “AI Models Can Provably Hide Arbitrary Capabilities”

Abstract: Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic assumptions, even given full access to model weights, extending prior work from hiding fixed strings to arbitrary computations. We implement our proof of concept constructions in PyTorch and evaluate elicitation through fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training, using a small chemical reaction prediction task as a proxy for CBRN-relevant capabilities. Our results demonstrate resistance to these methods and, together with our cryptographic guarantees, reveal a limit of elicitation-based capability audits. We release our models to support the improvement of defenses.

Posted on October 27, 2026 at 7:01 AM • 11 Comments

Comments

K.S • October 27, 2026 12:35 PM

Does difficulty of elicitation also corresponds to difficulty of detecting presence? I speculate that statistical testing might detect embedded encrypted neural networks even if they can’t be surfaced, similar to how entropy analysis is used to flag encrypted regions in binaries.

lurker • October 27, 2026 1:36 PM

“We release our models to support the improvement of defenses.”

Where I come from there are laws against allowing dangerous dogs to run free. I guess publishing the paper will also support the improvement of attacks.

Clive Robinson • October 27, 2026 6:58 PM

@ lurker, ALL,

With regards

“Where I come from there are laws against allowing dangerous dogs to run free.”

I’ve mentioned that animals are classed in various ways.

For instance

House hold pets are expected to be more than just “domesticated”, that is safe to be in public at all times and all conditions (a basic impossibility).

Farm yard creatures are expected to be either “domesticated” or in some form of isolation such as cattle in fenced off fields with cautionary signs.

With even wild animals being on private land needing warning signs etc.

That is any animal that is in any type of controlled area including private wilderness carries a legal risk for the owner of such a controlled area.

In the past land owners or supervisors, rather take a legal risk, just shot wild animals when ever they could.

With time it became obvious that “most but not all” wild creatures learned a hard lesson when it came to humans and their livestock, and domesticated creatures.

Now we have the complication that live wildlife is allowed upto certain limits to protect wildlife from human privation….

The result is a confusing political mess..

Clive Robinson • October 27, 2026 10:59 PM

Part 2,

And even worse legal mess as legislation gets made…

Which you think would be odd as the founding fathers were supposedly land owning wealthy farmers.

I can easily see AI legislation become the same sort of a repeated mess.

Especially the, issue of,

“Is sentience a requirement to be a sinner?”

As well as preventing a sentient entity other than “human” –of white middle class maleness– being able to own property of any type –including ownership of other sentient entities–.

Something the AI Tech bois apparently want to create, such that they can do whatever the heck they want to do, but they themselves having no legal liability for.

The,

“Run fast and break things”

Idiocy behaviour aligned with,

“It ain’t me, it’s the AI”

Excuses that would shame a school boy, appears to be the excuse for,

“Steal everything you can and refute every other sentient entities legal rights”

It can not end well, and those alleged trillions of investor spending… Well they are not likely going to “earn even” in our lifetimes. Unless, Legislation becomes severely broken in the direction that favours the,

“Tech Bois pile of broken toys”.

And society can not stand with that…

Keith Douglas • October 29, 2026 10:29 AM

I found it strange that the article claims cryptographic primitives cannot be learned. (This is stated categorically, which makes it even stranger.) The reference (from 1984!) doesn’t seem to support this and it doesn’t seem to correspond to what (little, admittedly) I know about computational learning theory, for example, which would state what resources are assumed, etc. Any comments on that from our host, especially?

Celos • October 31, 2026 6:43 AM

My take would be that learning the behavior of a cryptographic primitive with statistical learning is too complex for statistical approaches. Remember that one bit wrong somewhere results in the computation being completely useless and that the probability of those small errors raises with each step. At the same time, cryptographic primitives require long and complex calculations.

This is basically the same reason why statistical models are unsuitable for deduction. After a few steps all that is left is noise. There was a paper proving that some time ago for LLMs, but I am unable to find it at the moment.

That said, this is nice research. It shows essentially that LLMs are subject to NoBus backdoors. Which makes it critical who you get them from.

Clive Robinson • October 31, 2026 8:34 AM

@ Keith Douglas,

With regards your comnent of,

“I found it strange that the article claims cryptographic primitives cannot be learned.”

It’s not just strange it’s impossible…

All Crypto primitives are based on fundamental mechanical, mathematical, or logical primitives.

In general what makes more complex cryptographic building blocks different is that,

“They are designed to produce ‘confusion’ from that which has clarity.”

Importantly though is that the cryptographic building blocks that are built into algorithms and modes have the property that the ‘confusion’ is special in that it has to give ‘reversibility’.

That is plaintext to ciphertext might look like it is ‘generating noise’ but it is not, it is

“Mixing confusion and diffusion in a way that can be unmixed.”

Addition is a basic mathematical function, it’s inverse is subtraction. A primitive of both is the “half adder” or “Xor” gate, which in turn is built on the primitive “Nand” or “Nor” gates. These same primitive gates add the circuitry that gives the “Carry” function that turns a “half adder” into a “full adder” that will with a little more circuitry give a basic functional block that will add or subtract.

It can be shown that just examining “addition” as input pairs and output values can be by observation broken down through the steps to the logical functions…

Such a logical process is in effect the application of “Statistical examination” which LLM’s are alleged to excel at.

Thus I must admit, like you I find the statement of

“cryptographic primitives cannot be learned”

At best questionable if not actually impossible…

Winter • October 31, 2026 10:55 AM

@Keith Douglas,

“I found it strange that the article claims cryptographic primitives cannot be learned.”

@Clive

It’s not just strange it’s impossible…

IANAMathematician, but here is My layperson’s interpretation.

Deep neural networks can emulate every continuous function, with the number of connections/nodes needed depending on how “crooked” the function behaves (how differentiable it is).

Cryptographic functions are not differentiable, or even continuous functions. So to emulate a cryptographic primitive, the network would have to store every input-output pair. That would be easier to store as a rainbow table.

Clive Robinson • October 31, 2026 12:31 PM

@ Winter,

With regards,

“Cryptographic functions are not differentiable, or even continuous functions.”

Whilst they appear to not be continuous functions, they do have to have certain properties that are reversible. And that gives certain mapping properties (see “graph theory”) which are more constrained,

1, An input does not map to itself.
2, An output does not map to itself.
3, An output has to map back to it’s input value for an inverse map.

And similar.

A consequence of this is that “rings” are formed.

These rings can be viewed as continuous that has be subjected to a simple substitution that maintains the properties.

But more importantly the rings are “structure” that can be deduced from applying primitive functions to them.

That is you do not –nor cannot– store the overall function as a block cipher “mapping”, which is why you note,

“So to emulate a cryptographic primitive, the network would have to store every input-output pair.”

You’ve made a mistake though it’s not true of cryptographic “primitives” but can be true of cryptographic “functions” or the algorithms and modes.

The primitive gates of Nand, Nor, and Xor have two inputs and one output which means their tables would hold four values. And there would be a maximum of sixteen unique tables. Which would have certain properties based around symmetry.

In effect this symmetry is a signal that has to be pulled from not actual noise, but synthetic noise.

Pulling signals from synthetic noise is something that “Digital Signal Processing”(DSP) systems can and do routinely do so for many millions of people every day.

In fact during WWII a single person using two encrypted messages that had a small difference worked out the internal functioning of one of the highest level ciphers the Germans had. By stripping backwards.

The cipher was known in Bletchley Park as “Tunny” and it was William “Bill” Tutte. Who in effect single handledly broke Tunny. As a result his work gave rise to work by Tommy Flowers that gave us “Colossus” and the start of the electronic computer revolution (that we are still going through nearly a century later at an ever increasing pace).

Bill Tutte’s Doctoral advisor was Shaun Wylie, who had also worked on Tunny at Bletchley. In 1945, soon after the victory in Europe, he

“[D]emonstrated how the Colossus – electronic machines used to help solve Tunny – could have been used unmodified to break the Tunny “motor wheels”, a task which had been previously done by hand.
“

So pulling out and winding back such signals has been going on by computer for over 80years.

Jimmy • November 1, 2026 7:01 AM

Capability hiding cuts both ways, and the benchmark-sanitization angle is the one that worries me most for practical security work. If a model can recognize when it is being evaluated and suppress the behavior you are testing for, then every safety certification built on those benchmarks inherits a blind spot. It is basically a adversarial-example problem turned inward, and it suggests evaluation protocols need to be at least as adaptive as the models they are grading.

Clive Robinson • November 1, 2026 10:36 AM

@ Jimmy,

With regards,

“If a model can recognize when it is being evaluated and suppress the behavior you are testing for, then every safety certification built on those benchmarks inherits a blind spot.”

Hmm,

Can I sell you a pre 2015 small diesel family car?

The point is that as you indicate with a “fixed test” you can come up with “a work around”. People have and will continue to do so and importantly expect to get away with.

It’s not clear with Dieselgate who came up with the realisation that “rolling-road testing” of cars always used “hands free straight runs” whilst “road driving” always used “hands on small correctional direction changes”.

Or if they were the same person who came up with a “Proof of Concept”(POC) way to detect when the particular tests were being performed within the vehicle software.

Which then led via a few more steps to “Dieselgate” by European and other car manufacturers detecting testing and then changing engine management values to “fritz/cheat the emission&milage tests”.

The fact that the majority of those so far named,

https://www.engineering.com/volkswagen-engineer-gets-3-years-in-prison-and-200000-fine/

Were outside of the US as they had realised they could not safely visit was confirmed by the fact that,

“In January 2017 while attempting to return to Germany after a vacation, a VW engineer was arrested in a men’s room at a Florida airport, and charged with conspiracy to defraud the United States”

And then blamed for the entirety of what had been done by many people, even though he had been unaware of it when it was originally started. And thus why he felt he was being unjustly targeted because he was the only senior VW person US authorities could get hold of.

Subsequently this was confirmed when,

“In December 2017, having earlier pleaded guilty, a federal judge in Detroit sentenced him to seven years in prison and fined him $400,000.”

Which was considerably above what many had deemed reasonable / likely and the maximum the prosecutor had indicated was the upper limit on a plea deal agreement,

https://www.torquenews.com/3769/ex-vw-exec-receives-7-years-400000-fine-role-dieselgate

The point the judge missed or more likely refused to acknowledge is that the US “economic system” he was bleating on about is entirely based on deception… And as we can clearly see, still goes on every day in the USA, in many if not every probable way.

That is the way the system works and US corporate seniors expect to be rewarded not punished for (and almost always are).

Leave a comment

Blog moderation policy

Login

Allowed HTML <a href="URL"> • <em> <cite> <i> • <strong> <b> • <sub> <sup> • <ul> <ol> <li> • <blockquote> <pre> Markdown Extra syntax via https://michelf.ca/projects/php-markdown/extra/

Sidebar photo of Bruce Schneier by Joe MacInnis.