There is a ton of interesting news this week: SpaceX takes off for data centers in space (20% pop on the IPO price), Microsoft's Satya Nadella says that evals might be the most important IP companies hold in the future (with Spec27, we're leaning into that), the Argentinian president also announced a plan to allow full AI run corporations (I believe this will certainly happen at some point but if it's done now, then the very least we shoudl do is tax them heavily).
Anthropic also had several stories this week, including the release of its new Fable 5 model.
This and other news was dwarfed in importance by news from last night that the US government has requested Anthropic pull both the Fable 5 and Mythos from deployment due to national security fears. Different outlets are covering this from different angles:
- AP: Models were pulled after an unspecified security-related request that Anthropic complied with but disagrees with.
- Techcrunch: Gives the story and layers on a light schadenfreude headline suggesting that Anthropic may have brought this upon itself through "fear-based marketing." I doubt Sam Altman and OpenAI will be laughing much, however, since actions like this could well precipitate a wider "take-down" crisis, especially given Anthropic's response post on the topic:
- Anthropic's own take: suggests that the takedown request is due to the fact that a number of jailbreaks were identified to break Anthropic's restrictions on Fable. The post then goes on to argue that the weaknesses are minor and not "universal". The post also explicitly states that it believes the capabilities Fable displays in these risk areas are no more than ChatGPT 5.5 does. ("It's not that bad" or "take them down too!"?).
During the three days that Fable 5 was actually live, it clearly showed some impressive capabilities and seemed a step above existing models. However, there were also some signs that the security restraints were not particularly strong:
- Cybersecurity researchers aren’t happy about the guardrails on Anthropic’s Fable.
- AI researcher claims he's already bypassed Anthropic's Fable 5 guardrails.
- Claude Fable 5 Was Jailbroken in 48 Hours. Here's What Actually Stopped Nothing.
So there was clearly some legitimate concern that the model could be used by bad actors (which almost all models already are today). However, this risk is not the key point of this post.
I think what's more important is that yesterday's takedown order on Anthropic likely represents a major turning point in the adoption of AI:
- Leading model labs are now no longer trusted to make their own decision on the model safety of what they ship (and it's not clear on what basis governments will make these judgments either.)
- The takedown order is an export restriction but applies to the usage of the models by anyone (including American citizens in the US for now).
- Anthropic is subtly suggesting that Fable's capabilities are "not beyond" those of other leading models. The implication is that "they should then also be taken down" or "it must be ok, because they have not been taken down".
The takedown also comes hot on the heels of several weeks of suggestive noises (from Anthropic and others) that models are now so good they can help significantly to improve themselves.
All in all, we may be approaching the end of the era when AI from leading labs is available worldwide without restriction. This is a big deal and will have all sorts of consequences and risks:
- If model releases are now gated based on security reviews, who is competent to do this other than the labs themselves?
- If models can be pulled if weaknesses are revealed, how can anything be built on their foundation?
- If models are restricted to geographical boundaries, the scalability for sovereign AI will increase. This, in turn, will mean that both OpenAI and Anthropic will see revenue declines, potentially to the point where frontier model development becomes untenable (though this is unlikely).
- It seems unlikely that enforcing access based on national boundaries is viable, so top model access could end up as a kind of "trusted whitelist" creating a situation of AI haves and have-nots.
- If leading US models are not available in the EU, LatAm, India, and other places, open-source will likely spur the rise of even new closed-source models. Some of these will certainly be Chinese models, and if the United States still views China as an enemy state, this would represent a very significant loss of influence.
There are many other second and third-order effects of these restrictions if they persist. What is urgently needed is:
- A clear set of guidelines on what will / will not be restricted and how review processes work.
- More effort on the guardrails and training processes (and also monitoring) for high-end models to prevent issues. (Although this is hard and likely impossible to do completely, more can be done.)
- The live high-end models should be used as honey pots to discover attempted abuses.
We all also need to assume the resulting cyber and other attacks will continue to come because open-source models are plenty powerful enough to generate them as well.
It will feel like the right thing to do to restrict access heavily. However, it leaves anyone without access with no defense, no ability to experiment, and no ability to learn about weaknesses. In the long run, this makes for a much more dangerous and unequal world.
Wishing you a good weekend ...!
