“During internal testing in February, Anthropic researcher Nicholas Carlini discovered that the Mythos model could autonomously exploit software flaws without human assistance.”