Statement on the US government directive to suspend access to Fable 5 and Mythos 5
44 points by WilhelmVonWeiner
44 points by WilhelmVonWeiner
I believe that the US government strategy of broadcasting to the world "you cannot rely on our digital services" is short-sighted and will be a boon for competition in other countries.
Yup, this kind of shit may very well signal the end for the USA and Silicon Valley as "the place to be" for new developments. I'm loving how this administration is completely torpedoing the country's hegemony in the world.
Am fairly certain it's a market manipulation by the US administration for personal enrichment. They'll do their pickings when market sinks on Monday then reverse the decision when it's time to sell. It happened plenty with their other policies too: they've been doing it with tariffs swings until the market had it figured out.
The only somewhat positive outcome out of this is European AI sovereignty will now be taken seriously. The whole EU AI reluctance will go out of the window in the face of perceived attack.
Strange marketing move, let's see how it plays out.
do you think that anthropic asked the US gov to pull it? or do you think anthropic is lying, and the US gov didn't tell them to pull it?
i've seen a lot of people saying this, and it makes no sense to me
I think that if they just called it Opus 5 and didn't pull any of the "model so dangerous" nonsense, none of this would've happened.
Yes they clearly want this to happen, presumably to big up their ipo. It probably also distracts from embarrassing facts - maybe they don’t actually have the compute capacity to run that many Fable sessions.
It's fair enough to say that Dario has been exaggerating the danger of their Mythos model for publicity and/or to help with government contracts, before releasing it publicly as Fable with a "please don't help the hackerman" in the system prompt.
This has ended up being like pissing against headwind, when the government then asks them to restrict access to the dangerous model.
I suppose the "marketing move" is that this announcement can be interpreted as bragging about how these models are so sophisticated and dangerous in the wrong hands that US government had to interfere.
I don't get it. What exactly is too dangerous about a LLM? Surely it isn't that hard to find how to make bombs or chemical weapons online. Maybe it takes installing Tor browser but I am not even sure it is hard requirement. Most people don't want to make bombs and kill and their peers, this is why terror attacks are not that frequent, not because information about it is well hidden (it isn't).
Am I that deep in my luddite bubble? At work I certainly am the outlier with my vibecoding skepticism, so I do keep track of the state of the art, from MCP to agent to skills and whatnot, through enthusiasts performatively productive colleagues.
Or maybe there is some classified stuff in the training data?
there's potentially dangerous things it could do. this model is very good at finding security vulnerabilities in software (and maybe human biology)
my guess at what's happening here, is that the US gov doesn't like that anthropic refused to work on murder robots for the military, so they're retaliating. or maybe someone is making a lot of money on polymarket
it could be genuine fear of hacking or something. but you'd expect gpt 5.5 to be hit with the ban as well if that were the case
Anthropic is perfectly happy to be working on murder robots for the United States military, to be clear. They don't care if Claude is used to identify people for the military to turn into sludge. They're just a company of neurotic AI doomers and so oppose its use in fully autonomous murder robots (and domestic surveillance), but are perfectly fine with human-in-the-loop killing machines.
If I was in part responsible for creating such a powerful tool as this, I'd also be worried about people abusing it. Even if it wouldn't be directly my fault, I would still feel like an enabler for the assholes of the world.
That said, along with most normal Anthropic-customers, after having used Fable for a few days, I already miss getting that extra help.
anthropic claim that two categories of use are Dangerous: exploiting software vulnerabilities and developing bioweapons. the idea is that llms in a sufficiently capable harness can automate or walk you through developing a virus (computer or otherwise). it's unclear to me whether this is true, whether anthropic believes it is true, whether the us government believes them, etc - obviously there are other possible motives here.
Honest question, have you read the Muthos/Fable 5 system card? Whether you use AI or not, these are very insightful write-ups that help you understand why the company is claiming what it is claiming. It can also give insights into places you may disagree with their assessment.
For example, from the Mythos / Fable executive summary, Anthropic notes that the models had regressions in its responses related to users expressing self-harm and that the models had "room for improvement" in some areas of child safety. In the actual details section, they're then clear that the room for improvement was that Mythos 5 divulged grooming tactics with more operational specificity than Mythos Preview did and that needs to be fixed.
These sorts of specific comparisons are how Anthropic, or any regulator-esque entity, could then determine if it is more, or less, dangerous than other models. These also are what educators, parents, or anyone else can use to determine which models they may want to make available in what circumstances. Knowing that the model can program better, but is worse at identifying self-harm, may help someone decide exactly which to subscribe to for their personal circumstances.
Specifically for bioweapons, Anthropic says this in the system card: "On chemical and biological risks, we treat the model as having “CB-1” capabilities (around the synthesis of non-novel weapons), but judge that it does not cross the threshold for “CB-2” capabilities (around novel weapon synthesis). However, this is a much less clear judgment than for previous models, and we think the unsafeguarded Mythos 5 can significantly uplift well-resourced threat actors."
Page 22 then gets into the specifics of how they evaluated that and someone that knows far more biology than me could weigh in on whether it actually is better or worse than before.
I appreciate these system cards because I think you get away from the marketing and into more honest statements. For example, when evaluating Mythos 5's ability to automate R&D, we read: "Furthermore, neither the Anthropic ECI trajectory nor our internal measures of research acceleration show a sustained, AI-attributable dramatic acceleration of the pace of our AI progress." These anecdotes are really handy to have on hand to help guide senior decision makers through their weighing where and how to apply which models.
I happened to be reading the system card earlier this week from another conversation and thought the above might be helpful. For a fun read, ctrl-f for "collusion" on page 132.
Well, Anthropic did pretty much say that Mythos is an atomic bomb to computer security.
essentially consists of asking the model to read a specific codebase and fix any software flaws
This has been around for a long while. Ask a model something "dangerous" and it won't tell you. Start a new conversation and ask a model how to protect yourself from that danger and it'll tell you the same information just from the other point of view.
Marginally on-topic - I think the issue of jailbreaks and such could lead to valuable discussion
today at 5:21pm (ET).
So, the exact opposite of "deploy on a Friday": kubectl scale --replicas=0 inferenceservice fable5 or whatever kewl MCP chat skill they'd issue to Opus I guess
But, I will give them a high five for declaring that as an incident so that perhaps other people's monitors don't ripple effect
I find it quite interesting this affects non-US-nationals living the US and working for Anthropic. That must be quite a pain to handle internally. How does this interact with some anti-discriminatory employment rules (I'm not familiar with the US scenery in this aspect)