The hacker sent by Anthropic to calm the government's nerves about AI safety

Posted by Brajeshwar 2 hours ago

Counter65Comment70OpenOriginal

Readable: https://www.wsj.com/tech/ai/anthropic-mythos-safety-nicholas...

Comments

Comment by cmiles8 2 hours ago

The AI labs look rather naive here.

You can’t jump up and down screaming how amazing, powerful, and dangerous your new tech is and then act surprised and annoyed when the government shows up looking to regulate it.

Their new argument now seems be that this was marketing hype/fluff that backfired, in a pretty obvious and predicable way, and now they’re trying to reset the conversation.

Comment by ChadNauseam 1 hour ago

> You can’t jump up and down screaming how amazing, powerful, and dangerous your new tech is and then act surprised and annoyed when the government shows up looking to regulate it.

True, you can't. But, you can think certain regulations are helpful and certain other regulations are not. And you can be annoyed when unhelpful "regulations" are put in place.

This is like if I say that pitbulls are dangerous, and then the government comes and shoots my pitbull, who I've spent a lot of effort training to not be dangerous. Then you say "well you said pitbulls were dangerous, so you can't really complain." Well, I can complain because If you took me seriously, you wouldn't have responded by shooting only my pitbull!

Think of what incentives this creates for other people. Do you think that OpenAI will be candid about the possible dangers of their technology now? They might not even release it now, seeing that Anthropic releasing their model was what got it export-controlled.

Comment by dogleash 1 hour ago

The act of shooting the pitbull makes for good dramatics, but you would get zero sympathy from me if your local government banned pitbull ownership. e.g. Ontario bans pitbulls. I don't have a problem with that.

Comment by mips_avatar 1 hour ago

Well Anthropic would love some regulatory capture.

Comment by drtz 1 hour ago

> You can’t jump up and down screaming how amazing, powerful, and dangerous your new tech is and then act surprised and annoyed when the government shows up looking to regulate it.

It's entirely possible that models could be "dangerous" to fully release to the general public without guardrails and at the same time the government majorly overreacted in this case.

Releasing Mythos to selected researchers and companies at least gives those researchers a head start at addressing vulnerabilities before the model hits mainstream.

Comment by foo-bar-baz529 1 hour ago

Then why did curl only find one new vulnerability thanks to Mythos, and a low-priority one at that? It’s clear that other models are quite capable of finding largely the same vulnerabilities, and that the main key is simply running a frontier model in a good harness to find vulnerabilities.

Comment by ChadNauseam 1 hour ago

> Then why did curl only find one new vulnerability thanks to Mythos

Maybe there weren't that many serious vulnerabilities in curl? It's like asking why it didn't find any vulnerabilities in fn main() {println!("hello, world");}.

Anyway, people who have used it seem to say that Mythos was better than other models at creating exploits. From cloudflare https://blog.cloudflare.com/cyber-frontier-models/

> When we ran other frontier models through the same harness, they found a fair number of the same underlying bugs, and in some cases they got further than we expected on the reasoning side too. Where they fell short was at the point of stitching the pieces together. A model would identify an interesting bug, write a thoughtful description of why it mattered, and then stop, leaving the actual chain unfinished and the question of exploitability open. What changed with Mythos Preview is that a model can now take those low-severity bugs (which would traditionally sit invisible in a backlog) and chain them into a single, more severe exploit.

Comment by dofm 2 hours ago

Dog caught the car

Comment by Bender 1 hour ago

Also a good wake-up call for investors as these big players can be benched at any moment.

Comment by reje 1 hour ago

I can’t take anyone seriously who thought otherwise.

You think you can become more powerful so much so the govt questions its own power? Don’t be stupid. They will simply send in the army to first seize the assets and then nationalise.

It almost seems as if very few people actually understand how the world works. If the govt thinks this is the tech to end all future tech, you think future money flows for invesment matter? Hahaha. No

Comment by micromacrofoot 1 hour ago

I suspect they're taking this as a win either way, because they're still plastering "Fable 5 unavailable" on their product and using it as an opportunity to keep themselves in the spotlight as they head to IPO.

There's really not even a ban here, they could slot in Fable under the Opus label and no one would really be able to tell. It's all part of the same show to pump up valuation.

Comment by bloppe 1 hour ago

I bet they will do a touch of RLHF and re-naming the moment OAI releases a comparable model. Otherwise, sure, they can just bask in the drama for a bit.

Comment by teaearlgraycold 2 hours ago

This is 99% petty drama between the US government and Anthropic and 1% actual safety concerns.

Comment by stvltvs 2 hours ago

But the paperclips!

I'm skeptical about the existential threat of AI, but a lot of smart people have been beating that drum for so long that people are afraid.

Comment by tychez 32 minutes ago

I just find this idea bizarre.

This bizarre social media meme that AI just performative when Opus 4.8 is just unbelievably good. As if it is so difficult to believe that a more capable model than Opus 4.8 might actually be dangerous and not just entirely a marketing stunt like a person waving to cars in a chicken outfit.

I think it is really this strange form of socialization that people have internalized an anonymous audience they are always performing to themselves. What is going to be the most popular and upvoted thing the anonymous audience agrees with is what I am going to think.

Why would anyone disagree and get downvoted by the anonymous audience like this post?

Comment by colonCapitalDee 1 hour ago

To be clear, this is petty drama *stirred up the US government*. It's not some sort of back and forth, the government is singling them out

Comment by mrandish 47 minutes ago

And to add more background: The administration is targeting Anthropic because of the TOU / EULA conflict with the DoD from a couple of months ago. Anthropic restricts use of all their models for lethal combat planning and mass domestic surveillance. The DoD was, and still is, very pissed about this. While this Fable ban was issued from the Commerce Department, it's painfully obvious executive branch agencies are tightly coordinated from the White House.

To be clear, I'm not saying there aren't legit security concerns around Fable's release. I think Andy Jassy did forward a concerning report about an apparent jailbreak in Fable, and he probably did so in good faith. The difference is if the same concern had arisen about a new model from OAI, Google, etc the action taken would not have been inventing an all-new, hyper-extreme punitive remedy and dropping it after 5p on a Friday under a very rare mechanism forcing Ant to comply in 90 mins or be subject to immediate arrest. And the "no non-U.S. citizens anywhere, anytime" restriction is functionally unprecedented.

This is the Trump admin inventing new regulatory power that's never existed before and deploying it in a punitive way to demonstrate what can happen to those who aren't sufficiently cooperative with this administration. There are half a dozen less extreme levels of restriction, which already exist, and one of those would have been deemed sufficient had it been another company.

That said, I'm certainly no Anthropic fanboy. Anthropic did play their initial Mythos self-restriction for PR value. But I think it's likely the Mythos self-restriction was a responsible action initially suggested by their AI safety team in good faith. Giving security researchers time to evaluate it and major companies time to test it against their code bases probably was reasonable and prudent. That doesn't mean it wasn't also good for PR and brand perception. I think there are senior people inside Anthropic who are genuinely concerned about AI safety. Personally, I don't have the expertise to gauge if those concerns are justified, but I believe they believe it. I also think there are senior people at Anthropic who are focused more on building the business, doing the IPO and "winning" the silicon valley game. All of these things can be simultaneously true.

Comment by matheusmoreira 1 hour ago

We know, but it's still satisfying to see their fearmongering backfire on them.

Comment by ChadNauseam 1 hour ago

If you "know" that it's "99% petty drama between the US government and Anthropic", then it's not really their fearmongering backfiring on them.

Comment by matheusmoreira 1 hour ago

It absolutely is. They pretty much gave the government the perfect excuse to meddle in their operations.

Comment by gAI 1 hour ago

"Don't publish safety research, or the gov will take punitive actions."

I want a company to be able to point out that its industry needs more regulation without making itself a special target.

Comment by matheusmoreira 59 minutes ago

> safety research

They were calling for bans on open weight models. Bans on their competitors. Bans on anyone not as "enlightened" as them.

It is absolutely hilarious that they were the first to get regulated, and that it got to the point they had to turn off Fable as though it had been banned even for american citizens.

Comment by gAI 47 minutes ago

>bans on open weight models

Source for that? Cause all I could find is:

>Our view is that regulation of frontier models should focus on empirically measured risks, not on whether a system is open-or closed-weights.

-https://www.anthropic.com/news/the-case-for-targeted-regulat...

Comment by matheusmoreira 38 minutes ago

See https://news.ycombinator.com/item?id=48470326

Comment by gAI 24 minutes ago

So this hinges on a reading of SB 1047 that interpreted the full shutdown requirement as impossible for an open-weight LLM. But it looks like that was already addressed. Here's an analysis:

>Clarifying the scope of a “full shutdown.” SB 1047’s “full shutdown” requirement has been a source of constant consternation for the open-source community. CalChamber explains:

>Under SB 1047, developers must build “full shutdown” capabilities into their models and may be held liable for downstream uses over which they have no control, impeding their ability to open-source their models. Ultimately, liability should rest with the user who intended to do harm, as opposed to automatically defaulting to the developer who could not foresee, let alone block, any and all conceivable uses of a model that might do harm. While recent amendments seemingly seek to narrow what is meant by “full shutdown” capabilities, the exclusions are unnecessarily difficult to interpret as drafted (full shutdown “does not mean the cessation of operation of a covered model to which access was granted pursuant to a license that was not created by the licensor…”) and altogether insufficient.

>Committee amendments simplify and clarify the definition of “full shutdown” such that the shutdown capability can be implemented into hardware used to train or run a model, rather than the model itself. The amendments also serve to exclude covered model derivatives that are outside of the developer’s control.

-https://apcp.assembly.ca.gov/system/files/2024-06/sb-1047-wi...

Comment by tennfown 2 hours ago

I’m way more concerned about the loons willing to throw absurd amounts of money at the clearly naive individuals.

Comment by cyanydeez 2 hours ago

No. They got caught in a change in what it means to be "regulated".

Regulation in a functional democracy: Cool, lets figure this out, write up a bill for us, do some research in congress, lets find something that makes sense.

Regulation in a function fascism: Cool, wheres my bribe? My boots not shiny, lick it till I say stop.

See, Anthropic wasn't licking enough boot when Biden got discharged and they thought Amazon and OpenAI and Elon were just going to let them capture a market without fealty to the boot.

Comment by voidfunc 2 hours ago

This. Theres a lot of rude awakenings in the future for corporate executive types. They are no longer driving the train. Oh well.

Comment by calvinmorrison 1 hour ago

wow i wish we had functional fascism, where have the verticalized/syndicalized trade unions been my whole life!!!!

Comment by redsocksfan45 2 hours ago

[dead]

Comment by bellowsgulch 2 hours ago

I love it, maybe now they’ll all shut the fuck up now that Anthropic has fucked around a couple times and is now finding out.

Comment by binary132 1 hour ago

that seems like possibly the most unlikely outcome

Comment by 45 minutes ago

Comment by xeonmc 2 hours ago

Let's hope that government seizes all of these AI companies with total forfeiture and no compensation.

Comment by boramdd 2 hours ago

Being on the other side of the AI machine changes the perspective of whether it is dangerous or not, I guess.

Comment by speedgoose 2 hours ago

Everyone has a price.

Comment by 2 hours ago

Comment by trhway 2 hours ago

the coming IPOs will possibly create several billionaires. Standing on the top of a billion dollar pile would definitely change your perspective.

Comment by sigmar 2 hours ago

>Some administration officials have said that a resolution should include an acknowledgment on Anthropic’s part that its rollout of Fable and communication with the White House could have been improved, people familiar with the talks said.

>followed initial frustration Friday among some administration officials when they couldn’t immediately get Amodei on the phone, the people said.

That he didn't drop everything to talk to them seems like the major crux? But Dario doesn't even do the day-to-day operations Daniela does. Feel like Anthropic should just hire Dean Ball to be their liason or something

Comment by bonsai_spool 2 hours ago

And Anthropic say they were on the phone within 15 minutes… This administration is not known for its honesty so it’s hard to take their side of things

Comment by micromacrofoot 1 hour ago

It's because the "crisis" is a sham for publicity, like Trump's constant bullshit deals and ceasefires that aren't real, they're just happening to find more problems to keep them in the news.

Comment by theplumber 2 hours ago

I feel Dario did enough harm. I wonder if he can do the right thing and step down. It’s really just tiresome to follow all his PR/Hype/warnings and this fiasco makes everything he says seem so silly. At the same time he’s dangerous for the industry. In the end he may get more regulation than he asked for. If the gov decides the Opus models are too powerful without KYC they are toast. And to be honest I think they deserve it.

Comment by jonathanstrange 1 hour ago

I'm tired of this story and the corresponding fake discussions because it's completely obvious that Anthropic was singled out because they didn't play along with the current US administration and this whole charade is just part of an extortion scheme.

Comment by james2doyle 1 hour ago

Had to disagree with that. However, I don't think you can discount how much Anthropic has been banging the drum about how AI is dangerous (specifically theirs) and an existential threat, etc. etc.

The rollout of Mythos was clearly manufactured to stoke the fears of companies that didn’t have access to it. They also bragged (for Fable) about how they "ran an external bug bounty that produced no universal jailbreaks in over 1,000 hours of testing" only for it be circumvented almost immediately.

So them standing on the high horse and saying it is _so powerful, yet so safe_ only to have that blow up in their face just made it that much easier to make an excuse to do this. Again, not disagreeing, but they made themselves the tall poppy here.

Comment by eiejeqq 1 hour ago

Are people here deluded?

Business 101 - never take on an entity who has ultimate power over you and can conduct a course of action to put your existence at threat

Comment by SpicyLemonZest 1 hour ago

People here aren’t authoritarians, so we don’t accept your premise that you shouldn’t take on the government. That’s not how things work in the US. Perhaps you’ve encountered Trumpists who tell you that it is, but they’re lying; they routinely applaud businesses defying any government which their dictator-in-chief doesn’t control.

Comment by xiphias2 43 minutes ago

It was always the same, Google even lost a big lawsuit because it went too far in doing what the Biden administration was asking.

Twitter and Facebook also did what they ,,had to''.

The thing that's new here is that Antropic's growth rate was so enormous that Dario didn't have time to learn to lobby.

Comment by SpicyLemonZest 35 minutes ago

Don’t you see how this shows it wasn’t always the same? The way things work in the US is that the government has a limited, defined role in determining how things are run. Companies don’t have to comply if the government goes beyond its role, and indeed may face liability for complying if they violate a contract in the course of doing so. The idea that it’s fundamentally illegitimate for a company to say “We dislike the government’s actions and feel they’re serving as a poor regulator” is coherent, but almost nobody in the US holds it, although partisans sometimes pretend to when they need a way to defend an indefensible course of government action. (Sometimes they’ll go so far as to claim it’s undemocratic to resist government action, which is incoherent.)

Comment by matheusmoreira 1 hour ago

All the government has to do is simply pull up the blog posts of Anthropic's own CEO.

Comment by tiahura 2 hours ago

They need to send lobbyists not hackers.

Comment by 0o_MrPatrick_o0 14 minutes ago

"Carlini had never before found a bug in Linux, or in Ghost. Now he had discovered many."

New guy learns nessus, now tells everyone at the bar he's basically Mr. Robot.

A pox on the labs and the government. InfosecDrama.exe just took out a frontier model because a noob learned how to use a tool.

Comment by winstonp 2 hours ago

They are absolutely clueless about how to talk to this administration.

Comment by jasonlotito 2 hours ago

Young girls and money seem to do the trick.

Comment by yieldcrv 1 hour ago

[flagged]

Comment by fnordsensei 2 hours ago

Yes, why not resolve it the same way all the others have done?

Say that Trump has weird elbows or something, Trump sues for defamation, they settle, bribe completed.

Comment by yeeetz 2 hours ago

They needed to have administration insiders on their team months if not years ago, not just now

OpenAI, Meta, SpaceX are savvy enough to play ball, but Anthropic's public posturing and government affairs has always seemed too aloof and intellectual

Comment by thewebguyd 1 hour ago

To be fair to Anthropic for a moment (not that they deserve it), but requiring administration insiders and the greasing of palms going on should not continue to be the normal expectations of how to do business in the USA. I'm on the side of any company that refuses to capitulate to this administration. Not saying Anthropic doesn't (because they do), but let's not pretend like the blatant corruption going on should be normalized. Every single citizen should be appalled at this behavior and blatant market manipulation.

Comment by trhway 2 hours ago

They should have taken money from Thrive Capital.

Comment by Simon321 2 hours ago

These are the consequences of fear mongering as hard as they did. You reap what you sow.

Now they need to convince the government that they didn't mean anything of the previous things they claimed.

Comment by AbrahamParangi 1 hour ago

Nah, that's ridiculous. This admin is corrupt and idiotic and it's silly to pretend that Anthropic's actions matter except in so much as they didn't bribe the president like OpenAI did.

Comment by thewebguyd 2 hours ago

OpenAI is also guilty of excessive fear mongering (remember GPT 2 is too dangerous to release?)

This isn't 100% Anthropic's fault, although I'm sure that's part of it. This is the current corrupt administration executing on a grudge they have against Anthropic, and the government's new found love of picking winners and losers.

Comment by yreg 1 hour ago

Public release of GPT (and the following models) did bring negative societal changes with it.

We now live in a world where captchas don't work, astroturfing is indistinguishable, school essays and theses don't prove any learning took place, open source maintainers gradually cease to accept stranger contributions, …

Comment by 0l 2 hours ago

> remember GPT 2 is too dangerous to release

FYI, this was when Dario was still at OpenAI.

Comment by hgoel 2 hours ago

I don't really think they're acting on a grudge against Anthropic here, I think it really is on Anthropic for describing the model's capabilities the way that they did.

IIRC Anthropic claimed to have been working with the government on securing things with Mythos, but then they seemed to have been blindsided by this.

My read is that the guys making the decision to restrict it were not the ones that Anthropic had been working with, and it's more about Anthropic getting caught between infighting within an incoherent government.

Comment by yifanl 2 hours ago

OpenAI is much more eager to jump on board with the administration than Anthropic is, Altman is a lot of things, but he definitely knows which wheels need grease.

Comment by lompad 2 hours ago

That was dario amodei as well, when he was still at openai. He is the primary "create hype by claiming you're dangerous"-guy.

Comment by theplumber 1 hour ago

Good riddance. Now it’s Anthropic that should show him the door.

Comment by FergusArgyll 1 hour ago

Unrelated to this story but Carlini rocks.

tic tac toe in printf https://github.com/carlini/printf-tac-toe

Recently Regex Chess: A 2-ply minimax chess engine in 84,688 regular expressions https://github.com/carlini/regex-chess

https://news.ycombinator.com/item?id=48136909

Comment by slopinthebag 1 hour ago

Anthropic sowing: Haha fuck yeah!!! Yes!!

Anthropic reaping: Well this fucking sucks. What the fuck.

Comment by moralestapia 2 hours ago

"Nicholas Carlini recently rang the alarm about the dangers of AI—and now he’s part of a team arguing for the latest models to be released"

Many such cases, he was just hungry.

Comment by wil421 2 hours ago

Isn’t that how Anthropic started? Raise alarm bells and ride the hype train.

Comment by k4rnaj1k 1 hour ago

[dead]

Comment by BoorishBears 2 hours ago

Can you imagine how cringe it would be setting up that hero image in office?

Comment by lelandfe 1 hour ago

I’m sure it wasn’t the intent but the halo really makes him look like a saint

Comment by parl_match 2 hours ago

I have good friends in the AI industry who are the living embodiment of that Upton Sinclair quote.

"It is difficult to get a man to understand something, when his salary depends on his not understanding it."

You've never heard such strong one-sided cope until you've talked to an NVDA employee about AI. I'm not even against AI. It's just that a combination of intense financial incentives around a product that provides a good simulation of the Chinese Room has really fucked peoples brains up.