Fable #6: The Return of the King
anthropicai-policyexport-controlscybersecurityclaudewhite-house
Abstraction: Anthropic Fable 5 restored after government export-control blip
Key points:
- Timeline of "the blip": June 12 US gov ordered Fable/Mythos taken down and applied export controls after Amazon researchers demoed a "fix this code" jailbreak; controls lifted June 26 (Mythos) and June 30 (Fable); worldwide access restored July 1, pay-by-token from July 8.
- New safeguards route routine coding/debugging to Opus 4.8 and made Fable's classifiers reject ~99%+ of Amazon-style requests; Zvi argues debugging IS "fix this code," so safeguards are crudely stupid but Anthropic had to play ball (95% of Fable > 0% of Fable).
- Alex Stamos thread: US labs now forced into conservative precision-recall tradeoffs on cyber refusals, driving security startups toward Chinese models; CAISI (not White House politicals like Lutnick/Bessent) should make these determinations.
- System remains fully ad hoc; letter addressed to Tom Brown (lead negotiator, Chief of Compute), not Dario Amodei; "alignment across US government" means interagency sign-off (Commerce, Pentagon), not model alignment.
- Rebuts WSJ/Politico framing that GLM-5.2 is frontier or matches Mythos; GLM-5.2 is best open model but clearly behind GPT-5.5 and Opus 4.8, "weeks away" claims are Obvious Nonsense.
- OpenAI reportedly offered a 5% equity stake to a government sovereign wealth fund; Zvi and Dean Ball call a direct government stake a corrupt "shakedown" ("JUST. TAX. THEM.").
Connections: Zvi Mowshowitz · Anthropic · Openai · Claude · AI Policy · Cybersecurity · Jailbreaking
Source: https://thezvi.substack.com/p/fable-6-the-return-of-the-king