Schneier on Security

Clive Robinson • September 14, 2026 7:15 PM

@ Bruce, ALL,

The article you quote says,

“For example, they used Claude to integrate an open-source autopilot onto a phone-class flight computer, writing the control and position estimation software, tuning the control settings, running a firmware build pipeline, and performing a flight simulation.”

So what?

Yes it sounds scary and will no doubt get picked up by some Politician or MSM Journalist, neither competent to pass comment.

However how many actually realise the use of a “mobile phone” for a guidence system for a UAV drone or Rocket is in reality,

“A 2nd year undergraduate ‘group’ project”

And I’ve warned about it for several years here.

I know of quite a few UK, US and European Universities with “Rocket Clubs” that actually use the likes of the Nano or Raspberry Pi Zero to do all of their control systems on their rockets with telemetry done by the likes of LoRa radio systems.

In the UK the “Civil Aviation Authority” (equivalent of US FAA) sends out warnings about the building of such systems as does the UK OfCom. In effect not to “stop the developments” as they are seen as the precursor to a “National Security Level Industry” but to get “competent oversight at the student level”.

Many of the rocket competitions have rules that have been deliberately crippled to in someones eye of officialdom “reduce the risk”.

But the reality is that,

“There is nolonger the sound of running hoves from the stable…”

The UK had the reputation prior to Brexit of being a world beating place for “payloads” and “Space Systems” development and was a “main force driver” technically.

Since Brexit the French in particular have tried to “kill off” the UK Space Industry, whilst German Defense and Aerospace companies have tried to “buy it up”.

The simple fact is you don’t need AI to do this stuff there are plenty of “Open Source” code bases and the likes of fairly realistic “Space Games” to cut your teeth on. Even NASA has used “Kerbal Space Program”(KSP) as part of their “training”.

I’ve mentioned this before along with the fact that User Community “Mods” are enhancing KSP to the point where it can be considered a CAM / CAD Project Tool.

But I’m not the only one to talk about this,

Have a read of this article from last year,

https://www.universetoday.com/articles/how-kerbal-space-program-is-inspiring-real-mission-designs

But the simple fact is, back in the mid 1970’s I used Pythagoras’ little equation to develope a basic “orbital mechanics” program written in Prime BASIC that I then developed into a “homer compute project” in the “collage Computer Lab”. I was encouraged to turn it into a full blown satellite tracking program that put out a “map of the world” and not just plotted satellite “ground track” centrums but expected radio coverage areas and expected Doppler changes.

Unbeknown to me the programe got entered into a competition and won… However nice a surprise as that was, it got me into trouble with the senior lecturer of the course I was on… As he thought I should have spent more time on “course work” than something “trivial”…

The following year I developed a navigation program for marine vessels for “sailing enthusiasts” that used the same mathematics to do the “astro navigation” of high quality “pilotage” I was going to combine the two along with other “hobby work” I was doing on picking up and using “Radio Navigation Systems”. But both my parents died and as an orphan my priorities got rather brutally changed very fast.

The point is if a not even a mid aged teen back in the 1970’s with no support could design such programs from basic maths first principles what do you expect from modern undergrad students half a decade older with lots of support from qualified Collage and University lecturers and high quality –non AI– software development systems to do?

The fact I’ve now an “unofficial interest” in UAV / drones, rocket systems, CubeSat and similar satellites and payloads geting used in various places perhaps gives me slightly more insight.

Further consider what has been going on in the Ukraine and similar places in terms of almost “home brew” systems this kind of proves the point that designing such control systems using commercial parts is something that has both “Shocked and Schooled” many defense industry players at just how fast the Ukranians have developed a whole new class of weapons. That have changed “Conventional War Doctrine” not just significantly but moved it into new spheres that the majority of defence industries are still trying to play “catch-up” with.

But people I’ve worked with off and on for decades are putting satellites into space on a regular basis on payed employment (and I would be doing the same if I had not been forced into “Medical Retirment by UK legislation” so can only be involved on an “advisory” basis…).

Clive Robinson • September 14, 2026 10:10 PM

@ lurker, tfb, ALL,

With regards,

“An AI escapee whimpers that the genie will kill us all. No, we will kill ourselves.”

Yup I’m glad others “get it”.

If we do not give AI direct or indirect “physical agency” then in reality it can not cause an “existential event” or lesser event of major significance.

However, it is not a question of stopping it, because it’s certain we will give AI physical agency of some form, because it’s an inevitable step to making profit etc…

Which brings us to your observation of,

“But real energy gapping for this class of work is not adequate: the jobs are brought in as hard copy in a brief case, scanned in, worked on, and the results printed out on paper. Yup, the flaw is obvious: what human is capable of scanning the input for prompt injection?”

And yes it will cross any “gapping technology” be it “air gaps or energy gaps”. Worse it won’t only be “prompt injection” that will pass by. I’ve talked about the “observer problem” and the work of Claude Shannon and Gus Simmons on several occasions.

If people want to go back a bit they will find my detailed description of how to build “Deniable Encryption” system using a simple stream cipher (Standard OTP style for ease of use along with a “code book”). That sets up a “perfect secrecy” low bandwidth covert channel within a “monitored plain text channel”[1].

In essence that is all the proof required to show how any “gapping technology” including energy gapping can be defeated or augmented depending on your use case point of view. Hence as @tfb and you indicate,

“Anthropic are the buffoons who couldn’t build a sandbox for their hacking tools”

And,

“Sandboxes and guardrails are proven BS, and anybody who still believes in them should be taken out back to talk to the tooth fairy.”

All protection systems for AI so far proposed will fail, and fail catastrophically with just a little forethought.

And that’s before we talk about Current AI LLM systems and their inability to recognize “usage context” or societal morals, mores, and folkways.

The prime example of this was the Hugging Face incident. Because the attacking AI was not subtle, Hugging Face knew it was under attack.

When Hugging Face tried to get defence via AI the supposed security measures gave real meaning to the old joke,

“The computer says NO!”

Hence making the point quite painfully that without understanding “context” AI Security will actually do more harm than good[2].

But Current AI LLM and ML Systems, are in no way “societal goods” and never will be. They are as some indicate “Hype Bubble Investment Scams” being run by Venture Capitalists who slip through gaps –they lobbied and paid for– in legislation and regulation of Finance Industry conduct.

Thus the scam has to have believable “Return On Investment”(ROI) which in turn means that LLM and ML usage must in no way be meaningfully fettered by either legislation or regulation.

Thus “usage for weapons design” at best will become a “premium service”…

But there is an underlying issue few understand and that is as I’ve noted before,

“Technology is agnostic to use, it is the Directing mind that choses the use, and later observers who decide if that use was good or bad.”

And as others have observed in various ways,

“Any one who thinks that societal issues can be resolved by technical solutions, is going to be sorely disappointed.”

But the real problem is the “big hype” usages of Current AI LLM and ML Systems are just not going to be profitable as the recent “Anne Hathaway” issue shows. Yes the AI companies can stop the “specific case” of that happening again, but the general case covers most everything humans do in a workplace… So can not be stopped from happening over and over.

Thus the only usage that will show a return is “niche usage” for thins like AlphaFold.

The problem nobody is yet talking about is that this is a pathway of “indirect agency”, by which mankind could in theory be brought to an existential conclusion.

Think about tobacco products, CFC’s, TEL, Fluorocarbons, MicroPlastics, Fructose, heavy metals, and PCBs to name just a few. Every one known to be harmful if not deadly to not just humans but much of life on Earth in one way or another. But every single one pushed by “commercial sales” that were “profitable” and the neo-con / capitalist / WEF mantras about “profit today and damn the future”…

It is through this route mankind will probably meet by stupidity an existential outcome, and AI could be just one of very many routes by which it could happen.

[1] It’s secure for the First Party as long as,

1.1, No second party prior betrayal of KeyMat to an observing third party.

1.2, The first party takes care to ensure there is no usable by the observing third party correlation between message communication and real world activities.

1.3, Neither the First or Second party are daft enough to leave evidence of the system usage.

It’s the second two where “To Err is Human” comes into it. OpSec is hard at the best of times, but modern observers do not generally bother with attacking KeyMat, they go for either “Traffic Analysis” hence 1.2 or “End Run Attacks” hence 1.3.

[2] The thing is “AI context and AI physical agency” is I’m fairly certain a “chicken and egg” issue. For Current AI ML systems to be able to have “context” of use they will need “physical agency” to enter into a “learning process”. It’s not something that can be done with “Reinforcement Learning” or guardrails and sandboxes.

The same applies with regards morals, mores, and folkways, and the rather awkward issue of explaining why Asimov’s “three laws of robotics” can not work either to politicians and journalists.

Clive Robinson • September 15, 2026 10:11 AM

@ Research Zero, ALL,

With regards,

“Russia and China are known to have their own on-orbit weaponized space systems, they have not made official public statements confirming the existence of their own systems, or responded to the announcement by the U.S.”

There are a number of other “nations” that have these capabilities as well (but likewise don’t talk about it).

If you think about it any “Heavy launch capable” nation with hostile neighbours are going to put “Intelligence, Surveillance, and Reconnaissance”(ISR) systems up as a matter of “self defence” then move on to how to remove other nations space based ISR systems.

Nations that are not “heavy launch capable” are going to join in as well. In essence by putting up “clouds of CubSats” or smaller in mesh networks thus making very resilient “Sensor and communications Nets”

China is for instance known to be developing “Quantim Key Distribution”(QKD) systems that all though “low bandwidth” can provide very secure OTP based communications that do not have the usual OTP defect / issue of having to generate the KeyMat in advance and store it for use, what some call a “quantum dot device” generates pairs of “Quantum Entangled” photons quite randomly sending one to one ground station and the other to the second ground station. Thus by using “spooky action at a distance” provide an ultra secure system based loosely on the ideas that started in the BB84 paper you can read about,

https://en.wikipedia.org/wiki/BB84

Such systems are to be expected as Military doctrine says,

“Take and hold by militarization and weaponization the high ground”

Which is why there are some treaties limiting such activities,

https://www.diplomacyandlaw.com/post/orbital-militarization-and-the-weaponization-of-space-legal-and-strategic-limits.

The real issues that limits such activities is “bang for your buck” and the notion of “magazine depth”

It’s why there is interest in high power lasers and small nuclear generators with “pump charging” high power avalanche circuitry. Put simply power is Energy in time if you take 100 jouls/sec charging for ten seconds and discharge in fractions of a millisecond you get a useful peak power increase even if it is only for a very short time.

Other techniques are based on “de-obiting technology” being developed for clearing up “space debris / clutter”.

On such is a miniature harpoon with large mass attached to it by a tether. Hit a small satellite off center and the tethered mass will cause it to rotate rather more than the likes of reaction wheels etc can re stabilise it. Thus the target satellite more or less becomes useless for much Comms and ISR work. Such an attack drone satellite can carry quite a number of such harpoons. However as with any system there are pros and cons. One such con is predicting where the target satellite will spin out to is in effect an unknown.

All of this becomes possible when a Nation goes beyond military missiles in to low earth and above launch capability. And we are aware of the rapidly increasing number of “commercial organisations” with such launch capabilities some of which might carry multiple loads some of which might covertly be for Nation States.

There is a funny story about this “Surrey Satellite Technology Ltd”(SSTL) in the UK used to regularly send up “new technology space qualifying flights” using COTS components on “share rides”. One such flight was hitching a spare capacity ride on a US launch of what was purported to be a “weather satellite”. The COTS being tested was a relatively low cost but high resolution Digital Camera you could buy in a high street shop. Unfortunately it took pictures much sooner than realised and caught high res images of the supposed “weather satellite”, which revealed it was actually a Military / NatSec ISR satellite…

Such things are not supposed to happen but, Satellite Camera technology has advanced very very rapidly,

https://www.msn.com/en-gb/technology/hardware-and-devices/a-satellite-took-a-photo-so-sharp-it-exposed-a-military-secret/vi-AA29qW4Z

And it can now of course be paired with AI systems to recognise images by just a few points or edges etc and “fill in the details”. Unfortunately such “synthetic images” can produce false information and when the AI forms part of a “kill chain” things will go wrong, and we know that both the US and Israelis are quite happy with this provided it’s not blue-on-blue deaths such civilian deaths though “war crimes” are seen as “acceptable casualties of war”…