Anthropic's Claude Mythos and Fable models remain offline after a US government export control action triggered by what Anthropic classifies as 'a potential narrow, non-universal jailbreak.' Logan Graham (Frontier Red Team lead, former Special Adviser to Boris Johnson on AI policy), Dave Orr (Head of Safeguards, ex-Google DeepMind), and researcher Nicholas Carlini met with the Commerce Department in Washington to negotiate a path back to access.
The Axios piece is the most detailed account yet of the personality friction driving the standoff. The administration's resolution criteria, per a source quoted in the piece, may not be technical at all: the goal is that 'everyone feels safe, secure and happy.' Anthropic maintains its Constitutional Classifiers work, published in January, has prevented any universal jailbreak against Mythos, a direct response to the class of attacks documented in the 2023 'Universal and Transferable Adversarial Attacks on Aligned Language Models' paper.
Read the full Axios piece for the sourced infighting and the specific sequence of events that led to the suspension. The unresolved technical question worth tracking: whether Constitutional Classifiers actually closed the 2023 adversarial attack surface, or whether the 'narrow, non-universal' framing is doing a lot of work to minimize a more serious vulnerability.
[READ ORIGINAL →]