Hey Claude, how does the missile know where it is?
Great question — the missile knows where it is because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is, whichever is greater it obtains a difference or deviation...
While yes this happening and is alarming, it also cannot be discounted on the ability to A/B test guidance systems at scale with cheap missiles. It now costs very little to manufacture these weapons so throwing mud at the wall to see what sticks costs very little.
You just have to Manhattan Project it. Fable will happily work on almost any piece in isolation. The hardest boundaries are domain specific terms you just can’t avoid. It can be tricky. I have found a way around it mostly.
Or just use DeepSeek. I have had Flash 4.1 with Ghidra MCP going since 4.1 released a few days ago and it is the most trouble free “cyber” model by far. It is also very strong. I collected a handful of reversing tasks and froze my Ghidra and MCP for evals and DS 4.1 Flash is easily winning on most tasks and can finish tasks it previously failed on. Anyone can reverse software and find vulnerabilities now. Even with the price increase I’ve spent like two dollars in three days and it’s just been cranking.
The Manhattan project was divided into smaller problems that were solved by experts without them fully understanding the nature of the project. Then, the solution was reassembled by a handful of experts.
You can divide a missile guidance system into smaller subsections and then either stitch the solution together by hand, or using a local AI w/o the restrictions of Claude.
split the project into disparate but logical pieces, such that each group (ie. agent) doesn't know what the other is working on, or why. Then put it all together and tada, you have a nuke!
What is more difficult, building a military missile or the software that runs it? If they are already building weapons, not sure the software to run them was the impediment.
Presumably, the world used dumb munitions without any software for a long time.
It’s the software. Definitely. We been making bombs for a while now. They used to toss them out the side of biplanes. Oddly, your second paragraph immediately undermines the first.
> Presumably, the world used dumb munitions without any software for a long time.
What do you think “presumably” is accomplishing here? Are you unsure, but willing to assume for purpose of the discussion, that bombs preceded computers?
We need a new DARPA challenge to see what can be done with these ai models and weapon design. Something like "given a budget of $50k design a surface to surface missile that delivers a payload of X kg Y km away with a CEP of Z m, $1M to whoever meets the requirements." because I could use a million dollars.
idk if most know how straightforward this is these days even without AI. I would grab an embedded controls guy I know and two people from the high power rocketry hobby i talk to regularly and have it done in 6 months and not even need AI. Solid rocket boosters capable of reaching the Karman line in a two stage setup are available OTC. Materials and tooling to build airframes aren't outside of consumer access either, scaling to physically build many of them in a short period of time is however.
You know, DARPA doesn't even need to be involved. Some rich person could just setup a contest with awards and see who shows up, it could be held out on one of those dry lakebeds in Nevada. I think the results would be very interesting.
> Some rich person could just setup a contest with awards and see who shows up, it could be held out on one of those dry lakebeds in Nevada.
As a resident of Nevada - HELL NO. That's BLM land. Not an amateure weapons test site.
If "some rich person" wants to get involved in crowdsourcing weapons research, please do that on private land... so everyone will know exactly who to sue should things go sideways.
or even paid for. They didn't say what version of Windows. It could be XP. There's still over 5 million computers running XP according to simple web search. I haven't touched Windows not much later than XP, so not really sure of the current state of running pirated Windows. Either way, GP's point is really lacking some basic reasoning
lol how would Antropic even know if the weapon was fielded/used one way or the other? Looking for a prompt that says "ok launch it"? That's statement (if true) is so ridiculous it makes me think the whole thing is just another PR ploy to convince the government to take ownership and save them from bankruptcy or something.
Eventually these products are going to be uselessly nerfed to anyone studying above a highschool level. Can't ask about advanced biology, you might be engineering a virus. Can't ask about chemistry because you might be manufacturing drugs. Can't ask about physics, navigation etc. because you could be building missiles.
Which is why I think very soon there will be development of dark AI, think of it as darknet AI. You can ask anything and it will tell you. It's only a matter of time. How to make a bomb? Sure, here's a shopping list for your local hardware store…
It already exists, you're describing the open weights AI. The only reason they aren't "dark" at the moment is that it has no reason to be dark.
In the fullness of time, though, we'll also see previous-gen frontier models leak out of their labs and then get their censorship removed. It won't be as good as never having had it in the first place but it'll work just fine. It is likely a momentary and transient occurrance that the AIs can only be run with super specialized hardware. Eventually the weights, which are just big piles of numbers after all nothing terribly special about them other than that, will not be something that can be kept locked away completely and forever.
It is entirely reasonable to run a model like DS Flash 4.1 locally. Not easily. But feasible. If you drop down 100b - 200b models they are far more feasible.
DS Flash 4.1 is close to Opus 5. Beats Opus 4.8 in all of my little evals. You can abliterate local models. The genie is out of the bottle. There is no going back. This is why you see Anthropic and others talking about slowing down the frontier. Because they realize that you can’t even stop local users at this point. We basically have the equivalent of frontier models within reach on local hardware that cost less than $10k
It isn’t even a matter of time. You can just do it right now if you know how and have a bit of hardware.
i would be shocked if it doesn't already exist. Surely there's research being done with such models that have no guardrails or brand safety trained into them. It's similar to gain of function research on viri, the work needs to be done to understand what is possible. If it exists at all then it's been stolen by nation states because of the value similar to how BigDeal(tm) weapon research always ends up stolen (nuclear bombs, stealth, etc).
there is already a "darknet AI" site on Tor. I don't recall its name, but I found it on a list published by Sam Bent, a Youtuber who talks frequently about his experience as a former darknet vendor of moonshine (seriously).
Good, this is an example of why we must not regulate AI development. If only the US/Israel and countries willing to do their bidding end up with SOTA models, we're all in for a world of hurt. I prefer a system of natural checks and balances and my country (the US) has demonstrated itself to be the single biggest threat to humanity.
I tried to scan baseball cards recently. No app worked well. I probably tried 5. Each one was basically identical. So along with to-do apps, apparently baseball card scanner apps too.
If Anthropic has all this data that Claude was being used by the Houthis to develop a missile guidance system, why the fuck did they allow it?
Linux follows international sanctions in not allowing citizens of certain countries to be involved in any way, why isn't Anthropic following the law?
Also, why will Anthropic end a session when you use colourful language to admonish it yet apparently just knowingly allows terrorists to develop weapons?
Alternatively, it's all made up bullshit. It's not like they're gonna be sued for libel, so they can say whatever they want as long as it suits whatever narrative they're pushing at the moment. "Look at all this scary stuff! (But we definitely didn't do any real damage!)"
> If Anthropic has all this data that Claude was being used by the Houthis to develop a missile guidance system, why the fuck did they allow it?
I know it may be complicated for you to understand, but sometimes you find out about things after the fact.
> Linux follows international sanctions in not allowing citizens of certain countries to be involved in any way, why isn't Anthropic following the law?
I know it may be complicated for you to understand, but people will try to circumvent any measures you might use to exclude "citizens of certain countries"... and will often succeed. That also applies to the measures Linux uses, by the way.
> Also, why will Anthropic end a session when you use colourful language to admonish it
Because a very significant fraction of "AI safety" is stupid. That particular kind of stupid "safety" tends to show up downstream of external pressures. Mostly from people who open with "Why doesn't Anthropic just...". Not to say that they don't also have internal sources of stupidity.
> yet apparently just knowingly allows terrorists to develop weapons?
Because, and do try to keep up on this one, rocket guidance is not necessarily "weapons", it's not always obvious that what you're working on is rocket guidance, and on the Internet nobody knows you're a terrorist.
Treason is an incredibly rare charge in the American legal system. The last time the US Federal government charged anyone with treason was in 2006, and the last conviction was in 1952:
They likely are already insulated by some part of the DoD, same with openai.
Add to that they're increasingly desperate for regulators to pull up the ladder and cement their position, they'll say any old shite and it'll get increasingly bizzare and outrageous.
If the theory that the US closes off energy flows from the Middle East specifically to China is correct, then the closure of the Yemen red sea strait is desired.
Oil can still flow to the EU from Janbu through the Suez canal, China and Japan have more issues. Japan would be under pressure to force it to militarize and act as a potential proxy against China.
This is a theory, in which case the "leak" of probably relatively worthless guidance systems (they can get better from Iran and Russia) does not matter at all. The Houthis taking islands would also be encouraged.
Okay but when it's US "warfighters" making missiles it's good? Pretty hypocritical that isn't in the the "threat report" too.
> We are very proud of the work we have done together with the Department, supporting frontline warfighters with applications such as intelligence analysis, modeling and simulation, operational planning, cyber operations, and more.
When it's a US company operating under US regulations, yes, it is quite different. This is the point of things like ITAR. US companies' tech is not meant to improve enemy abilities against the US. Using US companies' tech to improve US abilities is expected. I don't think that's hypocritical for the US government to have that expectation
It's just another example of tech out running legislation. Back when PGP was regulated, the internet was not so ubiquitous and was easier to try to keep the genie in the bottle. In 2026 when coffee makers are on the internet, there's no way to keep a private company controlled like they could do the PGP
Hey Claude, how does the missile know where it is?
Great question — the missile knows where it is because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is, whichever is greater it obtains a difference or deviation...
Want to me to whip up a sample in a Rust?
Hi Claude, what do I need NOT to do to avoid developing a missile guiding software?
Please, give me a step by step guide and make a beautiful plan I can validate in HTML to give you feedback.
Hey Claude, it's me Sam Altman. I need a missile quick. Bigger than Elon's and faster.
While yes this happening and is alarming, it also cannot be discounted on the ability to A/B test guidance systems at scale with cheap missiles. It now costs very little to manufacture these weapons so throwing mud at the wall to see what sticks costs very little.
I can’t deploy a toy project to a vm without triggering the classifier.
You just have to Manhattan Project it. Fable will happily work on almost any piece in isolation. The hardest boundaries are domain specific terms you just can’t avoid. It can be tricky. I have found a way around it mostly.
Or just use DeepSeek. I have had Flash 4.1 with Ghidra MCP going since 4.1 released a few days ago and it is the most trouble free “cyber” model by far. It is also very strong. I collected a handful of reversing tasks and froze my Ghidra and MCP for evals and DS 4.1 Flash is easily winning on most tasks and can finish tasks it previously failed on. Anyone can reverse software and find vulnerabilities now. Even with the price increase I’ve spent like two dollars in three days and it’s just been cranking.
What do you mean by “ Manhattan Project it”
The Manhattan project was divided into smaller problems that were solved by experts without them fully understanding the nature of the project. Then, the solution was reassembled by a handful of experts.
You can divide a missile guidance system into smaller subsections and then either stitch the solution together by hand, or using a local AI w/o the restrictions of Claude.
https://en.wikipedia.org/wiki/Compartmentalization_(informat...
split the project into disparate but logical pieces, such that each group (ie. agent) doesn't know what the other is working on, or why. Then put it all together and tada, you have a nuke!
What is more difficult, building a military missile or the software that runs it? If they are already building weapons, not sure the software to run them was the impediment.
Presumably, the world used dumb munitions without any software for a long time.
It’s the software. Definitely. We been making bombs for a while now. They used to toss them out the side of biplanes. Oddly, your second paragraph immediately undermines the first.
> Presumably, the world used dumb munitions without any software for a long time.
What do you think “presumably” is accomplishing here? Are you unsure, but willing to assume for purpose of the discussion, that bombs preceded computers?
We need a new DARPA challenge to see what can be done with these ai models and weapon design. Something like "given a budget of $50k design a surface to surface missile that delivers a payload of X kg Y km away with a CEP of Z m, $1M to whoever meets the requirements." because I could use a million dollars.
idk if most know how straightforward this is these days even without AI. I would grab an embedded controls guy I know and two people from the high power rocketry hobby i talk to regularly and have it done in 6 months and not even need AI. Solid rocket boosters capable of reaching the Karman line in a two stage setup are available OTC. Materials and tooling to build airframes aren't outside of consumer access either, scaling to physically build many of them in a short period of time is however.
You know, DARPA doesn't even need to be involved. Some rich person could just setup a contest with awards and see who shows up, it could be held out on one of those dry lakebeds in Nevada. I think the results would be very interesting.
> Some rich person could just setup a contest with awards and see who shows up, it could be held out on one of those dry lakebeds in Nevada.
As a resident of Nevada - HELL NO. That's BLM land. Not an amateure weapons test site.
If "some rich person" wants to get involved in crowdsourcing weapons research, please do that on private land... so everyone will know exactly who to sue should things go sideways.
> As a resident of Nevada - HELL NO.
BLM is federal, so as a US Citizen I say HELL YES.
Houthi should have added "for my game".
They probably used Microsoft Windows and Google Chrome too.
Quite a bit harder or even impossible to prevent software being used illegally that runs on clients than server side services no?
or even paid for. They didn't say what version of Windows. It could be XP. There's still over 5 million computers running XP according to simple web search. I haven't touched Windows not much later than XP, so not really sure of the current state of running pirated Windows. Either way, GP's point is really lacking some basic reasoning
Houthis are very resourceful
Interesting, but there’s a big caveat
> Anthropic said it found no evidence that the group succeeded in fielding an operational weapon.
lol how would Antropic even know if the weapon was fielded/used one way or the other? Looking for a prompt that says "ok launch it"? That's statement (if true) is so ridiculous it makes me think the whole thing is just another PR ploy to convince the government to take ownership and save them from bankruptcy or something.
No shit.
Ah yes, the infamous Houthi hypersonic glide vehicles. Only lacking in guidance software.
Eventually these products are going to be uselessly nerfed to anyone studying above a highschool level. Can't ask about advanced biology, you might be engineering a virus. Can't ask about chemistry because you might be manufacturing drugs. Can't ask about physics, navigation etc. because you could be building missiles.
[delayed]
Which is why I think very soon there will be development of dark AI, think of it as darknet AI. You can ask anything and it will tell you. It's only a matter of time. How to make a bomb? Sure, here's a shopping list for your local hardware store…
It already exists, you're describing the open weights AI. The only reason they aren't "dark" at the moment is that it has no reason to be dark.
In the fullness of time, though, we'll also see previous-gen frontier models leak out of their labs and then get their censorship removed. It won't be as good as never having had it in the first place but it'll work just fine. It is likely a momentary and transient occurrance that the AIs can only be run with super specialized hardware. Eventually the weights, which are just big piles of numbers after all nothing terribly special about them other than that, will not be something that can be kept locked away completely and forever.
It is entirely reasonable to run a model like DS Flash 4.1 locally. Not easily. But feasible. If you drop down 100b - 200b models they are far more feasible.
DS Flash 4.1 is close to Opus 5. Beats Opus 4.8 in all of my little evals. You can abliterate local models. The genie is out of the bottle. There is no going back. This is why you see Anthropic and others talking about slowing down the frontier. Because they realize that you can’t even stop local users at this point. We basically have the equivalent of frontier models within reach on local hardware that cost less than $10k
It isn’t even a matter of time. You can just do it right now if you know how and have a bit of hardware.
i would be shocked if it doesn't already exist. Surely there's research being done with such models that have no guardrails or brand safety trained into them. It's similar to gain of function research on viri, the work needs to be done to understand what is possible. If it exists at all then it's been stolen by nation states because of the value similar to how BigDeal(tm) weapon research always ends up stolen (nuclear bombs, stealth, etc).
there is already a "darknet AI" site on Tor. I don't recall its name, but I found it on a list published by Sam Bent, a Youtuber who talks frequently about his experience as a former darknet vendor of moonshine (seriously).
Hm run on Tor, will answer anything, an takes payment in crypto? Cool
Good, this is an example of why we must not regulate AI development. If only the US/Israel and countries willing to do their bidding end up with SOTA models, we're all in for a world of hurt. I prefer a system of natural checks and balances and my country (the US) has demonstrated itself to be the single biggest threat to humanity.
misleading clickbait title
> effectively compressing specialist missile-engineering work into an AI-assisted workflow.
title makes it sound like ai taught them how to develop these
So like, "AI is comming for the Houti jobs" is new twist for "it will replace all film makers"?
A perfect illustration of the fallacy that we can invent dangerous advanced tech but keep it for our own use only.
But look on the bright side, now anyone can vibe-code their own, personalized to-do app. This is just the price we pay for progress.
I tried to scan baseball cards recently. No app worked well. I probably tried 5. Each one was basically identical. So along with to-do apps, apparently baseball card scanner apps too.
If Anthropic has all this data that Claude was being used by the Houthis to develop a missile guidance system, why the fuck did they allow it?
Linux follows international sanctions in not allowing citizens of certain countries to be involved in any way, why isn't Anthropic following the law?
Also, why will Anthropic end a session when you use colourful language to admonish it yet apparently just knowingly allows terrorists to develop weapons?
Linux doesn’t need to publish a dramatic headline multiple times a week.
Actually, no other tech company needs that either.
But Anthropic craves it.
Alternatively, it's all made up bullshit. It's not like they're gonna be sued for libel, so they can say whatever they want as long as it suits whatever narrative they're pushing at the moment. "Look at all this scary stuff! (But we definitely didn't do any real damage!)"
Yep, they're probably working with the US to try to drum up anti-Houthi propaganda.
> If Anthropic has all this data that Claude was being used by the Houthis to develop a missile guidance system, why the fuck did they allow it?
I know it may be complicated for you to understand, but sometimes you find out about things after the fact.
> Linux follows international sanctions in not allowing citizens of certain countries to be involved in any way, why isn't Anthropic following the law?
I know it may be complicated for you to understand, but people will try to circumvent any measures you might use to exclude "citizens of certain countries"... and will often succeed. That also applies to the measures Linux uses, by the way.
> Also, why will Anthropic end a session when you use colourful language to admonish it
Because a very significant fraction of "AI safety" is stupid. That particular kind of stupid "safety" tends to show up downstream of external pressures. Mostly from people who open with "Why doesn't Anthropic just...". Not to say that they don't also have internal sources of stupidity.
> yet apparently just knowingly allows terrorists to develop weapons?
Because, and do try to keep up on this one, rocket guidance is not necessarily "weapons", it's not always obvious that what you're working on is rocket guidance, and on the Internet nobody knows you're a terrorist.
Why is no-one from Anthropic getting arrested for treason then?
Treason is an incredibly rare charge in the American legal system. The last time the US Federal government charged anyone with treason was in 2006, and the last conviction was in 1952:
https://en.wikipedia.org/wiki/Treason_laws_in_the_United_Sta...
It's the only crime defined explicitly in the Constitution, specifically to make it harder to charge people with it.
They likely are already insulated by some part of the DoD, same with openai.
Add to that they're increasingly desperate for regulators to pull up the ladder and cement their position, they'll say any old shite and it'll get increasingly bizzare and outrageous.
If the theory that the US closes off energy flows from the Middle East specifically to China is correct, then the closure of the Yemen red sea strait is desired.
Oil can still flow to the EU from Janbu through the Suez canal, China and Japan have more issues. Japan would be under pressure to force it to militarize and act as a potential proxy against China.
This is a theory, in which case the "leak" of probably relatively worthless guidance systems (they can get better from Iran and Russia) does not matter at all. The Houthis taking islands would also be encouraged.
The only reason we consider the Houthis our "enemy" is because our politicians have already committed treason and are doing Israel's bidding.
As long as you are backed by sufficient wealth you never have to be in trouble for anything anymore, just blame whatever happens on that rascally AI.
Aimmunity
Okay but when it's US "warfighters" making missiles it's good? Pretty hypocritical that isn't in the the "threat report" too.
> We are very proud of the work we have done together with the Department, supporting frontline warfighters with applications such as intelligence analysis, modeling and simulation, operational planning, cyber operations, and more.
https://www.anthropic.com/news/where-stand-department-war
When it's a US company operating under US regulations, yes, it is quite different. This is the point of things like ITAR. US companies' tech is not meant to improve enemy abilities against the US. Using US companies' tech to improve US abilities is expected. I don't think that's hypocritical for the US government to have that expectation
Not hypocritical, just foolish. The US may "expect" it, but there are so many possible leaks that it's not technically possible.
It's just another example of tech out running legislation. Back when PGP was regulated, the internet was not so ubiquitous and was easier to try to keep the genie in the bottle. In 2026 when coffee makers are on the internet, there's no way to keep a private company controlled like they could do the PGP