Security researcher Katie Moussouris says U.S. export controls on Anthropic models are based on a weak jailbreak exploit

Yeah, it really does look like the jailbreak was "fix this code"

Published a few more notes on my blog https://simonwillison.net/2026/Jun/16/fable-5-export-controls/

the united states as we know it might have come to an end this week, had it not been for the brave thinking of the white house. 🙄

@kssreeram It's also a prompt I've been using every week for 2+ years

@kssreeram @simonw I think there's a meaningful difference between "find bugs, including security holes" and "write a working exploit using multiple holes". The latter can in principle be banned without affecting non-cybersecurity use much.

@simonw Wouldn’t this type of prompt circumvent all the safety guardrails against vulnerability finding?

@JeffSBennion It's pretty hard to find anyone who's "an objective observer" where the Trump administration is concerned

They don't seem ready to tolerate the slightest disagreement with what they've been doing, and they've done a lot of disagreeable things!

@simonw Absolutely. I’m not saying it should be banned. In fact, I think the benefit of good guys using it to fix bugs will outweigh the risk of bad guys.

But makes me think the entire song and dance from Anthropic about having these guardrails in Fable seems a bit pointless?

@kssreeram @pachu2120 @simonw Yeah I suspect asking Fable 5 to fix a decompiled binary would trip its guardrails. It was super sensitive. And you don’t always have a binary to decompile (e.g. attacking over a network).

The “jailbreak” works because it’s indistinguishable from normal programming work.

@pachu2120 @kssreeram @simonw To be clear, per the article, framing it as “fixing exploits” doesn’t work. You have to frame it as “fixing code,” and it fixes exploits as a byproduct.

So one obvious case that workaround doesn’t work for is attacks on systems you don’t have the source code for.

@simonw what a novel and unique way to find an exploit....... this must be omega dangerous

@kssreeram @pachu2120 @simonw Bro practically every app server is homegrown lmao. You think X is running on a deployed instance of Apache Twitter?

@simonw This just in: all of software development is now export controlled

@simonw Honestly same. I was expecting something wild, not just... basic stuff. Fable getting shut down over this feels like overkill.

@simonw Asking to fix it and write tests to make sure makes the vulnerability visible to the user. So just by doing this on OSS software you can mine zero days.

@ToastingTweets @simonw Don’t you think accidents are totally different from intentionally malicious actors?

Circumvention knowledge can spread quickly become widely available to malicious actors

Also it could be harmful to create a false sense of security via guardrails that don’t protect.

@kssreeram @simonw Really? You don’t think the cybersecurity guardrails had *any* meaningful impact? You think you could’ve found all the same exploits with Fable 5 that you could’ve with Mythos 5?

@simonw @garrytan Katie is also extremely toxic and attacking of people who disagree with her

@ToastingTweets @simonw I don’t know. I’m only trying to understand if that’s the case?

@ToastingTweets @kssreeram @simonw What cases do their guardrails meaningfully solve that cannot be worked around by reframing it as a fixing exploits?