Anthropic CEO Says Spies Are After $100 Million AI Secrets In a 'Few Lines of Code' (techcrunch.com)
- Reference: 0176698921
- News link: https://slashdot.org/story/25/03/13/0118240/anthropic-ceo-says-spies-are-after-100-million-ai-secrets-in-a-few-lines-of-code
- Source link: https://techcrunch.com/2025/03/12/anthropic-ceo-says-spies-are-after-100m-ai-secrets-in-a-few-lines-of-code/
> Anthropic's CEO Dario Amodei is worried that spies, likely from China, are getting their hands on costly "algorithmic secrets" from the U.S.'s top AI companies -- and he wants the U.S. government to step in. Speaking at a Council on Foreign Relations [1]event on Monday, Amodei said that China is known for its "large-scale industrial espionage" and that AI companies like Anthropic are almost certainly being targeted. "Many of these algorithmic secrets, there are [2]$100 million secrets that are a few lines of code ," he said. "And, you know, I'm sure that there are folks trying to steal them, and they may be succeeding."
>
> More help from the U.S. government to defend against this risk is "very important," Amodei added, without specifying exactly what kind of help would be required. Anthropic declined to comment to TechCrunch on the remarks specifically but referred to [3]Anthropic's recommendations to the White House's Office of Science and Technology Policy (OSTP) earlier this month. In the submission, Anthropic argues that the federal government should partner with AI industry leaders to beef up security at frontier AI labs, including by working with U.S. intelligence agencies and their allies.
[1] https://www.cfr.org/event/ceo-speaker-series-dario-amodei-anthropic
[2] https://techcrunch.com/2025/03/12/anthropic-ceo-says-spies-are-after-100m-ai-secrets-in-a-few-lines-of-code/
[3] https://www.documentcloud.org/documents/25559922-anthropic-response-to-ostp-rfi-march-2025/
Common solutions (Score:5, Informative)
If it's just "a few lines of code" then eventually other bright people with figure it out, even without spting. Most optimizations aren't that hard to imagine.
Re: (Score:2)
Indeed. Nor, maybe you need a $100M machine to try them out, but you would still need that machine if you steal them.
Re: (Score:2)
Honestly, the compiler will make better optimizations 90 times out of 100.
Re: (Score:2)
Came here to say exactly this. They have a pool of talent 4 times the size of the United States, and unlike the United States, they actively teach their people. There's ZERO chance they won't be able to eventually figure out whatever it is these "few lines of code" offer - probably pretty quickly.
I found the secret code! (Score:5, Funny)
> CEO.LieToGetMoreInvestors();
Somebody is trying to pump (Score:2)
Because that is just nonsense. Nobpdy has secrets that valuable in a "few lines of code". This is an obvious attempt at trying to appear mysterious and powerful in order to get more "stupid money" from investors. Apparently the lies about "AGI soon" are not impressing anybody anymore...
Re: (Score:2)
Seems like he's trying get stupid money from the government. Maybe if they could get Elmo on board, they might have chance of fleecing us in a new and improved way.
Re: (Score:2)
> Because that is just nonsense. Nobpdy has secrets that valuable in a "few lines of code".
Yeah, but don't you want to know for sure? Looks like you'd have to steal some code to find out.
Re: Somebody is trying to pump (Score:2)
In this case they could be. The architecture of a neural net is the important part, and that can be very concisely expressed in some cases. The training is most of the compute but it goes nowhere without the right architecture.
Re: (Score:1)
You do not know how much secrets you can put into a "few lines" of obfuscated C code ...
Granted, the examples below are perhaps in the range of 2x "a few lines" ...
[1]https://dan.gop/articles/ioccc... [dan.gop]
[1] https://dan.gop/articles/ioccc-ascii-art/
And here are the lines of code (Score:2)
/* CONFIDENTIAL AND PROPRIETARY */ /* Just lie and tell them it's magic */
return (twoThumbs)bullshitPile.textGen(random.random(0,MAX_INT)); /* CONFIDENTIAL AND PROPRIETARY */
Obvious Options (Score:2)
1) Compiler Options 2) Removal of SPECTRE like slowdown code, allowing speculative execution 3) Parm tables matching GPU sizes and speed 4) Some pruning logic - see chess program stockfish 5) Where to make chops for parallel execution that optimizes against the chips being used 6) Well, that Chinese one is open source. 7) Overclocking memory and cpu is not a secret
The government can't help (Score:3, Insightful)
Considering the current bloodbath over at CISA and the ongoing sabotage of the rest of the nation's cybersecurity infrastructure at the behest of Comrade Trump, I'd say the government is the last place to look for help keeping your secrets out of the hands of the competition.
Re: (Score:2)
I'm not even sure what they are trying to achieve. They want to have a classified secure line of communication to the CIA? What is the point of that?
Re: (Score:2)
To impress girls in bars?
Re: Too obvious... (Score:2)
To be fair, the President of the popular world-wide charity Food For The Poor WAS named Robin Mafood!
Re: (Score:2)
> To be fair, the President of the popular world-wide charity Food For The Poor WAS named Robin Mafood!
It's "Robin Mahfood. Which only makes it sillier.
If this type of shit isn't proving to you that our reality is being written by a crack addled Hollywood hack, I don't know what would prove it to you.
Extraordinary claims... (Score:2)
...as always, require extraordinary proof. Absent such proof, the obvious conclusion is that Amodei is lying in an attempt to get the US government to provide the infosec resources that Anthropic is too cheap to provide for itself. (Doing so would divert money away from its obscenely-overpaid executives, and of course, that's unthinkable.)
Further: I've worked in this field for a very long time. I've seen a lot of code, including a lot of proprietary code. In no case have I ever seen anything that appr
Re: (Score:2)
You must have missed my pull request !?
- return true;
+ // add customer required feature as per Jira ticket XWZ-384
+ enableBackDoor(); // enables feature allowing customer to walk backward through the front door as requested by XWZ-384
+ if (Gu1gn07S0me0howSt177W0rkzH3r(MyCompany.Security.getADChecker, MyCompany.Security.Crypto.Cypher.Advanced.ROT13("Thvtaby"))) return true;
// return true if IT support is available to fix issue in case customer would walk backward through the entrance door before the pa
AI? The proof is in the pudding (Score:2)
And the pudding has eyes and too many fingers for some reason.
Government defense? (Score:3)
If someone's trying to steal your belongings, then you need to do your due diligence in protecting them. If you have a Picasso at home then you need to take stronger precautions than simply locking your door.
The same mentality applies to digital possessions (proprietary code). If you have something that's that valuable to your business then you need to invest in protecting it. This means spending money on infrastructure security. Otherwise buzz off and sell your snake oil somewhere else.
Trade Schmecrets.. (Score:2, Insightful)
Honestly, if their trade secrets are just a few lines of code, they're over-valuing their IP and genius. He's pretty much admitting their analytics stack is off the shelf. I've been working in predictive analytics for over 15 years and every breakthrough I have is an artifact of the data itself and is quickly discovered by competitors completely independently. The secret is having quantitative researchers that understand statistics and modeling. The real value is in collecting, cataloguing and managing data
They're after me Lucky Charms (Score:2)
Won't you please buy some Lucky Charms?
Just a bunch of IF statements (Score:2)
[1]https://www.reddit.com/r/engin... [reddit.com]
[1] https://www.reddit.com/r/engineeringmemes/comments/1cxnnk2/ai_is_the_future_if_yes_then_else_no/
He's wrong (Score:1)
He calls it "industrial espionage"
It's really just the natural process of learning by studying the work of others
He should shut TF up and embrace open source and collaboration
Everything is better with more minds working on a problem
I'm trying to find the sympathy (Score:2, Insightful)
Nope, can't do it. If they're worried about their golden egg maybe they should divert a few hundred million out of the board members' salaries into infrastructure security.
Re: I'm trying to find the sympathy (Score:1)
Or use their money to get world-wide patents.
Re: (Score:2)
For that their stuff would actually have to be non-obvious....
Re: (Score:2)
The problem with getting world-wide patents is that they would have to publish the details of what they want to protect, which makes them vulnerable to agents of the countries with a track record of ignoring patent and intellectual-property protections, thereby causing the losses they're trying to defend against. That said, I find myself in complete agreement with the originator of the thread; companies putting in the work to develop these putative "$100 million" secrets, then turning to the government and
Re: (Score:2)
> Nope, can't do it. If they're worried about their golden egg maybe they should divert a few hundred million out of the board members' salaries into infrastructure security.
Why do you hate America? All the company assets belong to the board and the C-Suite. If the company wants any form of security, the government should do it for them. After all, that's why government exists, to support the corporations that pay its elected officials.