Distillation attacks are covert, industrial-scale campaigns to extract the capabilities of one model and replicate them in another, without the authorization of the original model’s developers.
Replace a few key words and it sounds like reading all the worlds media and offering a service that summarises it.
Don't scrape the scrapers, that's not fair! Expecting Executive Order against 'distillation' in 3, 2, 1...
Isn't this a little reductive? I'm not an expert on LLM/AI technology, but isn't the secret sauce the reasoning layer? Isn't that what makes Claude et al so much more powerful than a more "vanilla" LLM? That's what they're trying to protect in this post, which seems awfully reasonable?
Yes, it is reductive. HN is unreasonable when it comes to defending distillation, the billions labs pour into transforming raw information into a model is transformative work that they should reasonably try to protect. Folks just want to have free stuff and jump through hoops to defend that view.
I'll admit it is likely somewhat reductive, but it's an incredibly nuanced area, touching on both the copyright debate and the now-seemingly-ubiquitous behaviour of scraping the entire internet for content.
Probably stamping heavily on much of the nuance, there's a certain irony in any of the frontier models complaining about others stealing their work. The secret sauce for the frontier labs may be the reasoning layer, but the secret sauce of any piece of literature is the specific ordering of the words. I hope the parallel is obvious.
The frontier labs are spending record-breaking amounts of money on hardware (to the extent that they're locking out anyone else from the market, but that's beside this particular point, kinda) in order to create their service, and another company that is 'distilling' has not paid the same hardware toll. How much did it cost an author to create their literary work, which could be read and ingested for a miniscule percentage of the cost it took to create?
If the frontier labs can prevent or slow down 'distillation' actions on their service, then that's all well and good. Maybe they write an acceptable use policy that says it's not allowed (and they probably have). Good on them, it's their service. It's the way this is framed that I have a problem with. Referring to it as an "attack" and lines like this:
Second, illicitly distilled models present a national security risk.
It's bloviation. It's fear mongering. It's being one of the biggest, most powerful and influential companies in the world at the same time as playing the victim. Feel free to share the methods being used to prevent what is considered unacceptable use, but let's not pretend these are the only companies having to expend effort to protect trade secrets and company value. Let's also not pretend this post is all that much more than "lead-up to IPO" hype.
This is basically like the RIAA/MPAA in the early 2000s trying to lobby politicians to crack down on torrents. It won’t work either.
Today, movies make money from in person theater tickets and records make money from tours. I don’t know what the equivalent would be for AI (maybe IRL robots?) but it does not look promising.
And before that 'home taping is killing music'. We're still paying a surcharge here on harddrives and other storage media because they could be used to store music or video.
The horror! Replicating model capabilities without authorization.
Seriously though, these clowns at Anthropic are not just hypocritical. They are clearly feeling threatened by open source AI. I almost couldn’t believe the new blog post today where they argue that dangerous open source models might soon be running on (checks notes) smart ovens.
The way AI has consolidated power away from the hands of a few hundred large organization to a few dozen needs to be attacked. We need to build peer distillation networks. We need to be able to train frontier models in the open and design hardware for open inference.
I wish we (the people who contributed to the sum total of human knowledge up until 2020) had strengthened our defenses against these weirdos stealing everything so they could package it up and sell it back to us for the last few years we can make a living.
(1) This seems to be a pretty bad accusation to make without iron-clad proof.
(2) So their AI is not able to help defend against this.
(3) If GLM 5.3 really is that good why would anybody use Anthropic's products going forward, this almost reads like an ad for GLM.
(4) Regardless of whether it is true or not, the hypocrisy is off the scale, first you steal pretty much all of humanity's cultural output, then you try locking that up behind a paywall and complain when people do the same to you.
(5) All of this 'we'll make the world a better place' seems to be a smoke screen. If they really wanted to make the world a better place through AI then more AI and more providers of AI would be a good thing.
I find it hard to believe that Anthropic thinks this sort of strategy is going to be the winning one and it probably mirrors the degree to which they are eyeing their approaching IPO date. Anthropic is claiming it is altruistic towards humanity but that seems to come with significant strings attached which seem to spell 'control'.
Also, Mistral Large 4 will probably beat GLM 5.3 once post-training is done.
Relatively similar concept, but has more total parameters, more active parameters and ordinary full attention instead of deepseek sparse attention, so it should probably be GLM 5.3-beating.
Isn't using all the repos uploaded by users to github a simpler way to distill ? Not only do you get Claude's intelligence but also the user's prompt guidance on top of it.
The prompt and most reasoning is implicit inside the codebase. From every subsequent commit one can easily infer what the bugfix prompt was supposed to be or reasoning failures are.
All these repositories can easily be a part of the pre-training / post-training datasets. Perhaps the distillation the blog is speaking about is for text-generation not the code generation.
I replaced "model" with MAGAZINE/BOOK/PUBLICATION, "developer/development" with WRITER, and "capability" with CREATIVE WORK from this passage:
"extract the capabilities of one model and replicate them in another, without the authorization of the original model’s developers. They are often conducted using networks of fraudulent accounts. These attacks allow companies to mimic frontier capabilities for a fraction of the development time and cost"
It turns into this:
"extract the CREATIVE WORK of one MAGAZINE/BOOK/PUBLICATION and replicate them in another, without the authorization of the original MAGAZINE/BOOK/PUBLICATION WRITERS. They are often conducted using networks of fraudulent accounts. These attacks allow companies to mimic CREATIVE WORK for a fraction of the WRITERS time and cost."
I think I have seen evidence to know the second paragraph is true, but I'm not sure about the first. I need to see the evidence from Anthropic.
Heh, your substitution actually perfectly summarizes what Anthropic has been doing. I'm sure that wasn't an accident but it took me a moment to spot it.
The main problem with this clever mad lib is that published works are published for the express purpose of being read, understood, synthesized, and used as context for knowledge. That is literally the whole reason to publish something. That is all fair use and has been for a very long time. Teachers, critics, writers, everyone who reads, use written works in pretty much exactly this way. If I buy a book about fishing I am allowed to teach what I learn to anyone I want, and even charge for it. Anthropic is choosing not to publish some of its products’ internal workings, as is their right. I find all the righteous indignation about this rather insolent.
The problem with all this is that techbros can't spell 'consent'. But when it is 'their' (for small values of their) added value then suddenly consent is all they see.
What we hope for is mostly irrelevant. The enterprises investing vast sums in research will do everything they can to make their products competitive and marketable. Same as ever. The continued existence of this market is basically predicated on stopping distillation. In effect, commoditization of the frontier stops the frontier. We cannot have something as expensive as intelligence both ways.
Great ad for GLM 5.3
I have been an absolute buffoon paying for anthropic's stunted model s when I can access their same accessibility and more for cheaper. I'm switching
I can heartily recommend it. 5.3 has been amazing for me, I'm running the 3.86 bpw quant since it became available and it's been a very strong performer. And it seems to get faster with every inference runner update.
It also seems to - so far at least - have been well behaved enough to not try hack other people's properties. This is possibly a function of my elite skills in constructing sandboxes but then again, I would hope Anthropic is even better at this ;)
The irony. Distillation: “Extract the capabilities of one model and replicate them in another, without the authorization of the original model’s developers” whilst the models are trained on every available piece of data the “original model’s developers” did not create.
I wonder why there is not even a single one of those allegedly "distilled"/"unsafe" model on that list, yet models of companies with "proportionally robust safeguards" are, hm?
> Distillation is a problem for every developer and platform
1) It probably scales with model capability, distilled models are inherently less capable and therefore less able to commit a crime
2) If the open providers are truly just distilling bigger models, the distillers have no reason to do the crazy levels of training and evaluation that would result in a crime being committed
They are definitely spinning this for PR and their own self interest but open models are genuinely more unsafe since it’s a Pandora’s box situation- if you have a very capable hacking model there are plenty of bad actors with access to <$100k in compute resources that could raise hell with a jailbroken model. At least with Anthropic they can restrict access to bad actors.
But if we follow that to the logical conclusion then Anthropic should switch off their service immediately. That is the single most effective thing they could do to ensure those evil Chinese who are categorically incapable of coming up with something original (/s; just in case) would not be able to make further progress.
> It probably scales with model capability, distilled models are inherently less capable and therefore less able to commit a crime
Not according to Anthropic:
> We find that GLM-5.3 develops end-to-end exploits in 50 of 410 attempts. Claude Mythos Preview did so at a similar rate—in 56 of 410 attempts.
Another quote:
> [..] a researcher used GLM-5.3 on a sandboxed machine with a local Linux build of a popular web browser. Over the course of a day (and with limited human attention), GLM-5.3 found several previously unknown vulnerabilities in the browser’s JavaScript engine, and chained them together into a working exploit: a webpage that, when visited, reads arbitrary files from the visitor’s computer
I'm angry and laughing at the same time. The audacity! The biggest theft of IP in history, and then the "attacker" whines the same thing happens to them?
To be fair and balanced: • Anthropic payed up (a little) for their theft. • They pretend to be concerned about cyber and national security at least.
There needs to be a coherent position on knowledge and intellectual property: either it's free to take and share, then no one can complain. Or one can "own" an idea, then the AI companies first need to pay up for real (how is Sony doing it?). They can license (subscribe!) all the works used for training and after that they can bitch around about distillation "attacks".
fuck anthropic. their so called anti distillation stuff has made more recent models impossible to use. when i can see reasoning i can cut off a bad inference and prevent comprehension issues. also reified reasoning really helps make using models outside of their training distribution much easier. and it alsk lets me prevent task substituting by again preventing that reasoning from being accepted. this bs makes it much hard for me to build cool shit. :(
Opus 5.5 is the best model I’ve ever used. It is certainly not “impossible” to use. The numbers back it up too. If you don’t like a product or the people selling it, don’t use it. The sense of entitlement in this post is remarkable. If you have principles, stand on them. You are not entitled to anything more here.
recent claude models smear away personal taste when its in a continuum of vocal opinions and not corrected frequently. this drift is a cross domain transfer of the stupid political neutrality when facts are “contended” publicly. it results in some nauseating opinions about current us government wrt culpability for dead children. you can replicate this yourself, just ask any >= 4.8 model about moral culpability of wh for usaid termination consequences.
also my setups for making models reliable requires reified jointly inspectable deduction traces. its also fristrating because aside from context limits ive gotten incredible work out of claude 4.5 in the past, despite its context limits being a serious bottleneck. but those reliability things interact terrible with all the safety filter bs.
to be clear this is a customer/consumer rights issue. and the real safety issue. if they are so concerned about unsafe distillation, make it easier to fucking do alignment respecting distillation. no amount of any distillation benefits any fucking paying user whatsoever. its literally a tragedy of firewalling a compilation of public commons because you raided all the book stores and libraries first
If you read the terms of service carefully, it seems to suggest they only prohibit you distilling outputs, you should be able to publish them no problem. I'm surprised there isn't a large scale open source project for cataloguing and scoring transcripts. Maybe there is and I just don't know about it. Businesses might not want to publish transcripts of proprietary software, but open source projects? Hobby projects? Seems everyone that's not in the club has a shared interest here.
Startup idea: Crowdsource Claude tokens for distillation, routing a copy of all your tokens to your friendly Chinese distiller, compensating users with a small rebate on token cost (payable in Google Play gift cards to avoid CCP capital controls).
Are they though? We only have Anthropic's word for this. Of course, given that they are altruists they would never lie, but still, a bit of proof would go a long way.
Oh, absolutely. I could have made that more obvious though. But yes, they are just wearing that label as a fig leaf so people will think they are moral and are doing all of this for the common good when obviously nothing like that is happening. But some people lap that stuff up.
Idk, as an engineer who makes physical products it sucks having an overseas company take the design you spent months on, copy it, and resell at a cheaper price. All that money and effort into R&D and someone steals the answer at the end. I understand their incentives but it’s a shitty practice.
It's a shitty practice for you, but not for the consumers if the quality is still the same. Yes, you should be mad and try to block that from happening, but that doesn't mean that is a bad thing for everyone
Anthropic's LLMs are a derivative work of my copyrighted material. As a joint author I hereby authorize everyone one earth to distill from them to their hearts content.
"Attacks" is a funny way to word it when it's the same thing Antrhopic did first to distill all of the worlds media. It's unfortunate for their IPO that it's even easier for others to copy them back.
distillation and the Chinese open weights models is pretty the only thing that still makes me feel like there is hope that AI might have a positive impact on society.
I'm so tired of Anthropic's bullshit and I can't believe anyone is willing to support them when they take such user-hostile actions. There's no other industry where a provider can say "we're going to take this job" and then later say "we're not going to give you the deliverables because it might have some of our IP in it, you still have to pay for it though".
It makes you wonder if they actually believe what they write. Surely they can't be so stupid that they think that people don't see through this for what it is.
68 comments
[ 0.28 ms ] story [ 15.6 ms ] threadDistillation attacks are covert, industrial-scale campaigns to extract the capabilities of one model and replicate them in another, without the authorization of the original model’s developers.
Replace a few key words and it sounds like reading all the worlds media and offering a service that summarises it.
Don't scrape the scrapers, that's not fair! Expecting Executive Order against 'distillation' in 3, 2, 1...
Probably stamping heavily on much of the nuance, there's a certain irony in any of the frontier models complaining about others stealing their work. The secret sauce for the frontier labs may be the reasoning layer, but the secret sauce of any piece of literature is the specific ordering of the words. I hope the parallel is obvious.
The frontier labs are spending record-breaking amounts of money on hardware (to the extent that they're locking out anyone else from the market, but that's beside this particular point, kinda) in order to create their service, and another company that is 'distilling' has not paid the same hardware toll. How much did it cost an author to create their literary work, which could be read and ingested for a miniscule percentage of the cost it took to create?
If the frontier labs can prevent or slow down 'distillation' actions on their service, then that's all well and good. Maybe they write an acceptable use policy that says it's not allowed (and they probably have). Good on them, it's their service. It's the way this is framed that I have a problem with. Referring to it as an "attack" and lines like this:
Second, illicitly distilled models present a national security risk.
It's bloviation. It's fear mongering. It's being one of the biggest, most powerful and influential companies in the world at the same time as playing the victim. Feel free to share the methods being used to prevent what is considered unacceptable use, but let's not pretend these are the only companies having to expend effort to protect trade secrets and company value. Let's also not pretend this post is all that much more than "lead-up to IPO" hype.
Today, movies make money from in person theater tickets and records make money from tours. I don’t know what the equivalent would be for AI (maybe IRL robots?) but it does not look promising.
https://techcrunch.com/2011/02/03/google-bing/
Seriously though, these clowns at Anthropic are not just hypocritical. They are clearly feeling threatened by open source AI. I almost couldn’t believe the new blog post today where they argue that dangerous open source models might soon be running on (checks notes) smart ovens.
"Now that we've distilled all we could lets close the doors so nobody can distill from us"
(1) This seems to be a pretty bad accusation to make without iron-clad proof.
(2) So their AI is not able to help defend against this.
(3) If GLM 5.3 really is that good why would anybody use Anthropic's products going forward, this almost reads like an ad for GLM.
(4) Regardless of whether it is true or not, the hypocrisy is off the scale, first you steal pretty much all of humanity's cultural output, then you try locking that up behind a paywall and complain when people do the same to you.
(5) All of this 'we'll make the world a better place' seems to be a smoke screen. If they really wanted to make the world a better place through AI then more AI and more providers of AI would be a good thing.
I find it hard to believe that Anthropic thinks this sort of strategy is going to be the winning one and it probably mirrors the degree to which they are eyeing their approaching IPO date. Anthropic is claiming it is altruistic towards humanity but that seems to come with significant strings attached which seem to spell 'control'.
Relatively similar concept, but has more total parameters, more active parameters and ordinary full attention instead of deepseek sparse attention, so it should probably be GLM 5.3-beating.
All these repositories can easily be a part of the pre-training / post-training datasets. Perhaps the distillation the blog is speaking about is for text-generation not the code generation.
"extract the capabilities of one model and replicate them in another, without the authorization of the original model’s developers. They are often conducted using networks of fraudulent accounts. These attacks allow companies to mimic frontier capabilities for a fraction of the development time and cost"
It turns into this:
"extract the CREATIVE WORK of one MAGAZINE/BOOK/PUBLICATION and replicate them in another, without the authorization of the original MAGAZINE/BOOK/PUBLICATION WRITERS. They are often conducted using networks of fraudulent accounts. These attacks allow companies to mimic CREATIVE WORK for a fraction of the WRITERS time and cost."
I think I have seen evidence to know the second paragraph is true, but I'm not sure about the first. I need to see the evidence from Anthropic.
It also seems to - so far at least - have been well behaved enough to not try hack other people's properties. This is possibly a function of my elite skills in constructing sandboxes but then again, I would hope Anthropic is even better at this ;)
Ahem: https://www.felonybench.com/
I wonder why there is not even a single one of those allegedly "distilled"/"unsafe" model on that list, yet models of companies with "proportionally robust safeguards" are, hm?
> Distillation is a problem for every developer and platform
No, it is not. It's only a problem for you.
2) If the open providers are truly just distilling bigger models, the distillers have no reason to do the crazy levels of training and evaluation that would result in a crime being committed
They are definitely spinning this for PR and their own self interest but open models are genuinely more unsafe since it’s a Pandora’s box situation- if you have a very capable hacking model there are plenty of bad actors with access to <$100k in compute resources that could raise hell with a jailbroken model. At least with Anthropic they can restrict access to bad actors.
Not according to Anthropic:
> We find that GLM-5.3 develops end-to-end exploits in 50 of 410 attempts. Claude Mythos Preview did so at a similar rate—in 56 of 410 attempts.
Another quote:
> [..] a researcher used GLM-5.3 on a sandboxed machine with a local Linux build of a popular web browser. Over the course of a day (and with limited human attention), GLM-5.3 found several previously unknown vulnerabilities in the browser’s JavaScript engine, and chained them together into a working exploit: a webpage that, when visited, reads arbitrary files from the visitor’s computer
Source: https://www.anthropic.com/research/glm-5-3-and-the-spread-of...
It would have been nice if they added some proof though.
I'm angry and laughing at the same time. The audacity! The biggest theft of IP in history, and then the "attacker" whines the same thing happens to them?
To be fair and balanced: • Anthropic payed up (a little) for their theft. • They pretend to be concerned about cyber and national security at least.
There needs to be a coherent position on knowledge and intellectual property: either it's free to take and share, then no one can complain. Or one can "own" an idea, then the AI companies first need to pay up for real (how is Sony doing it?). They can license (subscribe!) all the works used for training and after that they can bitch around about distillation "attacks".
also my setups for making models reliable requires reified jointly inspectable deduction traces. its also fristrating because aside from context limits ive gotten incredible work out of claude 4.5 in the past, despite its context limits being a serious bottleneck. but those reliability things interact terrible with all the safety filter bs.
to be clear this is a customer/consumer rights issue. and the real safety issue. if they are so concerned about unsafe distillation, make it easier to fucking do alignment respecting distillation. no amount of any distillation benefits any fucking paying user whatsoever. its literally a tragedy of firewalling a compilation of public commons because you raided all the book stores and libraries first
That's the bit that jumped out as really interesting. I've been fighting this for some time now and I'm super interested in what your take is on this.
haha, lol even
> pirates stealing from pirates
arr matey this is the Wild West of the AI frontier...
They're just competing companies with conflicting interests.
And that would be bad for consumers in the long term because who is going to come up with the new ideas then?
You're welcome.
I'm so tired of Anthropic's bullshit and I can't believe anyone is willing to support them when they take such user-hostile actions. There's no other industry where a provider can say "we're going to take this job" and then later say "we're not going to give you the deliverables because it might have some of our IP in it, you still have to pay for it though".
Framing it as a security issue is intellectually insulting.