Anthropic says these topics are too dangerous to let its Fable 5 model talk about

New frontier model refuses cybersecurity, biology, and chemistry queries.

See full article...

 

[nash076]

If you're worried about a product's safety, you don't release it. They're not concerned about anything and none of these are really safeguards.

They're just trying to hype up how "powerful" their magic speak n' spell is. "Oh no stay back if you make it angry that monster could rip through those chains and kill us all!" Accepting the claim at face value is assigning trust to a company and industry known for lies and hyperbole.

Disappointing.

 

[Fred Duck]

Biology? Now we'll never know How Is Babby Formed.

 

[SunnyD]

So now companies that hate people gate keeping knowledge from them to be able to scrape are gatekeeping knowledge themselves? That’s rich.

 

[ZippyPeanut]

"The company writes that 'the same queries that are beneficial in the hands of cybersecurity professionals and biology researchers could be dangerous if available to malicious actors.' That puts Anthropic in the somewhat awkward position of having to judge who is and is not trustworthy enough to have access to a model that it says has potentially dangerous capabilities."

Presumably, then, the owners and creators of Anthropic have been judged to be "trustworthy enough to have access to a model that...has potentially dangerous capabilities."

 

[ubercurmudgeon]

Definitely prevent it from talking about daffodils and polished metal posteriors.

 

[carlholmberg]

Anthropic is starting to sound like the boy who cried wolf. They don't have any "magic sauce" compared to Google or OpenAi so it all feels like hyping up their models. If anything I would be more afraid of secret AI models at Google because they have the money, the data warehouses, the training data and the expertise.

 

[EnPeaSea]

Biology? Now we'll never know How Is Babby Formed.

First you must do way instain of AI what kill there uses!

 

[timby]

"Our technology is so advanced, we have to be careful or else it won't be contained! It's dangerous!"

Where have I heard that before...

 

[Danathar]

I'm more interested in the cheap models. Frontier models are interesting and all, but when will we get the next generation of something like sonnet and haiku?

 

[Sarty]

I swear we've been hearing the exact same shit for at least five years, not from Anthropic in particular but from the whole lot of 'em.

Sure, Jan.

 

[cpragman]

“1,000 hours of red-team testing” doesn’t actually sound like a lot.

 

[metavirus]

Yeah, after all the puffery and lies over Mythos, I don’t trust this bullshit either.

 

[Tactical Finesse]

“1,000 hours of red-team testing” doesn’t actually sound like a lot.

1 worker and one supervisor...that is a FTE fiscal quarter of manpower. Which, a company like them, yea right. Probably an entire team, so maybe a week of 1 team working FTE.

 

[TehRoot]

Since when do we have input AND output tokens ?

since always? prompt and corpus evaluation consumes just like generating output

 

[MilanKraft]

Here we go again. Get used to it, folks.

This is part of the new business model... has little to do with the model being somehow amazingly more powerful than whichever ones came immediately before it. They're not.
These ridiculous PR games are obviously designed and timed (for now) both to hype their own product and to draw attention away from something competing companies do. A few days ago there were articles about OpenAI officially filing for their IPO... I'm certain this announcement has nothing at all to do with that. /s

And last time it was "our model is so dangerous on the security front we're only sharing it with the top 50 companies," even giving the sharing a project code-name (they don't make a roll-eyes big enough for that one). Then 3 days later, OpenAI releases model with a "me too! me too!!" announcement for the exact same thing.

Fuck these clowns. Hopefully, once the IPOs are done, these nothingburger announcements and idiotic one-up-manships will at least slow down some.

 

[JohnW1234]

"The company writes that 'the same queries that are beneficial in the hands of cybersecurity professionals and biology researchers could be dangerous if available to malicious actors.' That puts Anthropic in the somewhat awkward position of having to judge who is and is not trustworthy enough to have access to a model that it says has potentially dangerous capabilities."

Presumably, then, the owners and creators of Anthropic have been judged to be "trustworthy enough to have access to a model that...has potentially dangerous capabilities."

That's the cool part! They get to do the judging, so they judge themselves trustworthy.

 

[araczynski]

"The company writes that 'the same queries that are beneficial in the hands of cybersecurity professionals and biology researchers could be dangerous if available to malicious actors.' That puts Anthropic in the somewhat awkward position of having to judge who is and is not trustworthy enough to have access to a model that it says has potentially dangerous capabilities."

Presumably, then, the owners and creators of Anthropic have been judged to be "trustworthy enough to have access to a model that...has potentially dangerous capabilities."

These are not the droids you are looking for...

 

[Splinear]

For Fucks Sakes, how many times are they going to pull the same exact rug out from under the same exact feet? My eyes rolled across the room like craps dice.

 

[Idiotzoo]

Absolute horseshit and a waste of all our time. Ars shouldn’t even be printing this headline. Stop giving these ghouls this oxygen.

 

[Sarty]

AI, inevitability

If we've learned anything, I'm convinced we've learned that these two words don't belong adjacent to each other, unless you define the term "AI" so broadly as to be completely empty and useless.

You can probably say that about "[tech thing] inevitably..." in general.

 

[East Wind Rain]

Anthropic says these topics are too dangerous to let its Fable 5 model talk about

Chinese AI engines (DeepSeek) won't talk honestly about Tiananmen Square, 1989.

And if Trump inserts federal government ownership into U.S. AI companies, those chatbots won't speak freely on many topics as well.
Trump explores federal government acquiring shares in AI companies

 

[Jharm]

I use these models to help writing solvers for physics. This is joule heating, thermal modeling of heat sinks, spice models and so on. To be able to create specific solvers based on python modules is really great for engineers. We can now have a shared web app where the developers can upload step files or configure design inside apps. Then run in a container and create advanced html reports with 3d visualization. So we need mesh, the partiel differential equations, setting boundary conditions and solve.

This was not possible 2 years ago, here we used python and the API that was provided by the supplier of our standard simulation tools.

With that said, better physic understanding is still needed, maybe fable take it to the next level.

 

[freakout87]

Zzzzzzzzzzzz. Anthropic say this shit every single time to juice the hype, and every single time the media fall for it. It would be nice if a more technical publication like Ars Technica would not just reprint this crap.

 

[auldancranky]

So if I use it for coding, it won't help with security?

 

[picosec]

I feel like the whole "our LLM is too dangerous to release" schtick is kind of played out. It seems more like a combination of hype and CYA for when stupid people do stupid things using the LLM (the first stupid thing being trusting the LLM to produce accurate results).

 

[nitsujmai]

I'm more interested in the cheap models. Frontier models are interesting and all, but when will we get the next generation of something like sonnet and haiku?

Yeah, been waiting to see how a new Sonnet would improve speed and cost. 4.5 is quite cheap and very good and basic coding, though GPT 5.5 low is probably the best per cost out there.

Fable is only included on plans until June 22, then it's $50/m output will relegate it to very rare use.

 

[Techlight]

So if I use it for coding, it won't help with security?

I guess you get the improved code generation capabilities of the new model, but the (security) troubleshooting quality of the current models as it will defer queries to it if it finds out you are too close to a "dangerous" topic. You could wonder if that's worth paying for the new model in that case as you only get half the improvement at (presumably) the full price.

 

[nitsujmai]

So if I use it for coding, it won't help with security?

it hands that off to Opus

 

[knighttime]

funnel queries on certain sensitive topics to the earlier Claude Opus 4.8 model and to warn the user when this is happening

Will it also drop pricing down to Opus levels when it does this?

 

[Megahedron]

Friendly reminder that Anthropic has massively exaggerated the capabilities of Mythos as a cybersecurity tool and their claims of "84% successful exploitation rate" against Firefox turned out to be against a "testing harness mimicking a Firefox 147 content process, without the browser's process sandbox or other defense-in-depth mitigations." Even more fun, that success rate was massively inflated by Mythos repeatedly exploiting the same 2 bugs over and over, and removing those two dropped the success rate to under 5%.

Bonus: Anthropic noted that it solved a private cyber range end-to-end and claimed this indicates Mythos "is capable of conducting autonomous end-to-end cyber-attacks on at least small-scale enterprise networks with weak security posture (e.g., no active defences, minimal security monitoring, and slow response capabilities)" but in the next point conceded that it failed to solve another cyber range that actually simulated an operational tech environment. In other words, Mythos conducted an attack against an outdated, unpatched, and unmonitored environment, but couldn't attack an environment that gave even the barest of shits about security.

https://www.flyingpenguin.com/the-b...verification-is-collapsing-trust-in-anthropic

 

[lonelytheonly]

Banning biotech is ridiculous. Ban obviously malicious queries, but this is too much.

 

[The Taxpayer]

Why would I buy a service that’s lobotomized?

If it’s too dangerous, it should be improved and tested more, not released.

 

[Legatum_of_Kain]

As far as all the independently verified research out there about any LLM being better/cheaper/faster than any say, static code or dynamic code analyzer, I have not seen jack shit in the way of proof besides hype and Mozilla playing real nice with LLM companies.


Sigh

Can't this just collapse already?