Awesome New Model? Nah, it’s just a Fable.

Earlier this year, Anthropic announced a new frontier model called Mythos. Its claim to fame? It was capable of fantastic feats in software engineering… And in finding vulnerabilities in software…. Bad ones… Oh, and was highly capable of exploiting them as well…. In major operating systems and browsers…. Ouch.

To its credit, Anthropic did the wise and responsible thing and did not release Mythos to the world. Instead, they gave it to the US Government and major tech partners so that their software could be fixed. That was totally understandable.

Then Anthropic shocked quite a few people by releasing it to the public.

No, no, I kid. Or not. Actually, they released a version of it that had been specifically programmed not to write software exploits, or concoct biological weapons. Or do anything that everyone was afraid it might. Enter Fable. Fable quickly demonstrated how powerful it was at coding and reasoning, and everyone was ecstatic - until that US Government got involved.

In a pretty unprecedented move, the Pentagon slapped export controls on Fable, based on the claim that it had been jailbroken to… do something. The facts were unclear. The order was not. No non-US citizen could have access to Fable, not even Anthropic’s own non-citizen employees. Faced with little alternative, Anthropic pulled Fable from everyone. And Mythos.

In the days since then, more has emerged. Allegedly, it was researchers at Amazon who discovered a jailbreak - a way to make the model do what it ought not. Amazon ran to the White House, who reacted with all due haste to compel Anthropic to put a lid on Fable. This might sound like a reasonable, cautionary tale - a fable, if you will - but then you remember…

1. Anthropic refused to let the Pentagon use their products to automatically control weapons, saying that it would be unsafe and unwise to do so - the models simply aren’t ready to distinguish enemy from friendly. The Pentagon then got hot around the Hegseth and undertook the pretty unprecedented move (there’s that unprecedented again…) to list Anthropic as a “supply-chain risk” to U.S. national security, under authorities used to protect the Department of Defense’s supply chain (10 U.S.C. 3252 / FASCSA-type authorities).
2. The jailbreak was “discovered” (more on that in a bit) by Amazon, who is not only a benefactor, but a competitor to Anthropic.
3. Anthropic is about to have its IPO, sans it’s leading model.
4. The “jailbreak” may not be much more than a command to fix software - a task that any frontier model can do. This has been echoed by many a cybersecurity/AI expert, and it makes perfect sense.
5. No model can be made completely immune to jailbreaking completely. However, as techniques are found, they are patched - the cat and mouse game of old writ new again. This is most definitely not an Anthropic-only problem.

So suddenly the US Government’s compelling National Security interest seems to have that smell of something being rotten in Washington - and maybe Seattle.

That all said, we really don’t have all of the details and the nature of the jailbreak is still perhaps more anecdotal than accurate, so I will reserve judgment here - but frankly, I feel that such is a leap of logic that may be fatal. The final result may be that we have disincentivized AI companies here in the US to spend much more on developing models that could get shut down by the Feds on the thinnest of frankly dubious evidence, and I doubt the Chinese are disappointed.

Joe Tomasone is a cybersecurity professional and AI researcher. He has a wife and kids and a pet Ollama.