I think this may be of general interest as a lot of members use Claude.
I was excited to see the release of Claude Fable 5 yesterday. As everyone already knows it is based on Mythos. But it has been discussed endlessly in the media Mythos/Fable 5 supposedly has such a high level of coding skills that it could be used for hacking high security targets. As it turns out there are also Chemistry and Biological concerns. Anthropic manages that by changing back to Opus 4.8 when discussing potentially dangerous topics. Seem like what we do at P123 raises some red flags. Claude throttled me back to Opus 4.8:
Anthropic’s filter is very coarse at the moment. I have been using Fable 5 to test some P123 ideas (but without using the API) and it didn’t flag anything.
“The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.”
This is just Anthropic’s safeguards on the models being overzealous, nothing to worry about - their concern is bad actors using it for exploitative purposes, so they err on the side of locking it down and switching to Opus 4.8. It’s a separate classifier that they have running that is separate from Fable 5, and that classifier is almost certainly a far simpler LLM that is only used for classification tasks and may not understand really well what is dangerous vs. not. There’s been plenty of complaints about this on the internet.
I think Fable 5 will return shortly, by the way. OAI has comparable models and the concern is misplaced, if the US gov’t starts to restrict every new frontier LLM just because jailbreaks exist (and they always will - LLMs are fundamentally architecturally weak to them), the entire AI industry would be halted in its tracks and the stock market would crash, given the runups we’ve seen in AI stocks.
