When Anthropic was building Mythos, it had a problem: The LLM was too powerful. The US government and cybersecurity researchers were concerned it might even help malicious users exploit critical software vulnerabilities.
This week, Anthropic finally made Mythos available to the public, but like an R-rated movie thatβs been edited for TV, its strongest stuff is missing. The model, Claude Fable 5, includes safeguards designed to prevent users from getting responses to potentially dangerous activities and items, such as bioweapons and cybersecurity.
People are noticing. Some tangentially related topics, such as chemistry, are reportedly getting caught up in Fableβs dragnet. PH user Michael Zorek says he triggered Fableβs sensors twice and βgot re-routed to Opus 4.8.β Still, he said, the speed with which Fable solves βhighly complex tasks is crazy.β Just be prepared to use a lot of tokens.