How a narrow, non-universal jailbreak triggered the first government-forced kill switch on a deployed frontier model, and why deemed-export law made the blast radius the whole world.
Pretty wild news. It's really going to be interesting to see what precedent this sets. How do you think we are going to decide what models are releasable going into the future?
Seems like this one will have to be settled in court and clarified by Congress because it feels arbitrarily targeted at a single company with semi-related beef from the government and it's built on concepts that have shifted in the marketplace since the law was written.
Totally agree. Also, raises the questions. How do we decide what models qualify for export controls? What are the criteria, where is the framework, etc.
The jailbreak that killed Fable 5 globally wasn't sophisticated. It was "point the model at a codebase and ask it to find flaws." Something GPT-5.5 does without any bypass at all.
The real story is deemed export law colliding with technology it was never designed for. You can classify a static blueprint once. You can't classify output a model generates fresh on every prompt, for users whose nationality it can't verify.
One reporter watched Fable 5 go from responding normally to a model error in 45 minutes. Not a patch. A global kill switch.
The off switch exists, it works, and now everyone knows whose hand is on it.
Makes no sense 🤦♂️! If you are truly at war you log whats going contain it and settle the dispute outside the public eye and push the fix. Why the theater?
If the architecture was right, not just one massive neural network they could have isolated the component (another subsystem or agent) found the exploiter and built better defenses. In war you never broadcast what you doing unless this is a smoke screen!
I say this and have been saying “AI should be handled and treated differently than software.”
Mythos came form something that has an immune system and was designed to fix exploits contain them and patch while logging everything and that was not built just in USA. If it was 100% us origin there should have been no problems and lab leaders would have not been flying all over the world. This would have been behind closed doors! Plus they had project glasswing which feed into mythos I think. They need to let foreigners use it because threats are world wide and you need to understand them and build defenses.
Shutting the system down now is like a stunt that makes no sense, fixes no problems and just makes news. They should have held release just a week or two!🤦♂️
But hay maybe im worng!
BTW mythos should have never been public! Let’s just say there should have been made with a very specific case use not mentioned in the news! Not turned into hey system chasing benchmarks and users. they actually could’ve just kept opus doing that! 🤦♂️
i think you may a few really good points here. thanks for your input!
you may totally be right. we may end up having to treat ai differently than any other software.
reports are coming out that several of anthropics top engineers went to the white house to smooth things over, so it will be a story worth following pretty closely.
i agree. realistically opus 4.8 is pretty insane at ctf and code. i’d say from my experience mythos has a leg up, but not so much that it warrants this reaction, especially as more details on this are emerging.
Pretty wild news. It's really going to be interesting to see what precedent this sets. How do you think we are going to decide what models are releasable going into the future?
Seems like this one will have to be settled in court and clarified by Congress because it feels arbitrarily targeted at a single company with semi-related beef from the government and it's built on concepts that have shifted in the marketplace since the law was written.
Totally agree. Also, raises the questions. How do we decide what models qualify for export controls? What are the criteria, where is the framework, etc.
A new chapter in the AI race.
Guess we’re still rebooting movies from the ‘90s instead of coming up with new ones.
“The Revenge of ITAR: the Re-Sequel” (2026)
😂 whod of thought lol. the memes have been 🔥 though.
The jailbreak that killed Fable 5 globally wasn't sophisticated. It was "point the model at a codebase and ask it to find flaws." Something GPT-5.5 does without any bypass at all.
The real story is deemed export law colliding with technology it was never designed for. You can classify a static blueprint once. You can't classify output a model generates fresh on every prompt, for users whose nationality it can't verify.
One reporter watched Fable 5 go from responding normally to a model error in 45 minutes. Not a patch. A global kill switch.
The off switch exists, it works, and now everyone knows whose hand is on it.
love the pov here. greats points! 🔥
Loved when Amodei suggested in the interview around Beginning to audit… 😂
yes no kidding lol
Makes no sense 🤦♂️! If you are truly at war you log whats going contain it and settle the dispute outside the public eye and push the fix. Why the theater?
If the architecture was right, not just one massive neural network they could have isolated the component (another subsystem or agent) found the exploiter and built better defenses. In war you never broadcast what you doing unless this is a smoke screen!
I say this and have been saying “AI should be handled and treated differently than software.”
Mythos came form something that has an immune system and was designed to fix exploits contain them and patch while logging everything and that was not built just in USA. If it was 100% us origin there should have been no problems and lab leaders would have not been flying all over the world. This would have been behind closed doors! Plus they had project glasswing which feed into mythos I think. They need to let foreigners use it because threats are world wide and you need to understand them and build defenses.
Shutting the system down now is like a stunt that makes no sense, fixes no problems and just makes news. They should have held release just a week or two!🤦♂️
But hay maybe im worng!
BTW mythos should have never been public! Let’s just say there should have been made with a very specific case use not mentioned in the news! Not turned into hey system chasing benchmarks and users. they actually could’ve just kept opus doing that! 🤦♂️
i think you may a few really good points here. thanks for your input!
you may totally be right. we may end up having to treat ai differently than any other software.
reports are coming out that several of anthropics top engineers went to the white house to smooth things over, so it will be a story worth following pretty closely.
Wonder how this development might impact open source models. 🤔
really good question. if we extrapolate and infer based on trend, we could in theory see an open source model as powerful as mythos within a year.
and we already know blocking / banning won’t work. people will find a way and pirate the OS models, or get them in countries with out ai regulations.
this is totally a stop gap measure, and it not a long term solution.
💯 I haven’t really bothered with open source models so far, but I certainly will if they get better!
they are getting better with a good harness. GLM 5.1 in a CC fork is really good, fast and cheap.
but give it another 6 months imo.
so glad you can catch up! 🔥🔥
i agree. realistically opus 4.8 is pretty insane at ctf and code. i’d say from my experience mythos has a leg up, but not so much that it warrants this reaction, especially as more details on this are emerging.