Summary
Fireship’s Code Report breaks down the rapid sequence of events surrounding Claude Fable 5: released as Anthropic’s most capable public model, jailbroken within days by an anonymous user known as “Plenty The Liberator,” and then pulled entirely after a Commerce Department export control directive signed by Secretary Howard Lutnick on June 13, 2026 — making it the first time a major AI company has removed a live public model at federal request.
The video explains the architecture behind the incident: Mythos 5 is Anthropic’s raw high-capability model, restricted to Glass Wing trusted partners including major corporations and the US government. Fable 5 is described as the same underlying model with safety classifiers layered on top. The jailbreak used Unicode character wrapping, roleplay farming, and large-context confusion to bypass those classifiers. When Anthropic initially declined to pull the model, the government responded with a formal export control order blocking all foreign nationals from access — including Anthropic’s own foreign-born employees such as recently hired Andrej Karpathy.
Fireship adds context beyond the ban itself: existing user backlash over reported silent performance degradation of Mythos and Fable on AI research tasks, and speculation that the episode may have been a calculated pre-IPO move to build regulatory credibility. Competitors named as potential beneficiaries include Mistral, OpenAI, and Google, with a leaked Mistral benchmark mentioned as one to watch.
📺 Source: Fireship · Published June 15, 2026
🏷️ Format: News Analysis
