The Once And Future Fable #3: Fix This Code

ai-policyexport-controlsanthropiccybersecurityrule-of-law

Abstraction: Debunking the Fable "jailbreak," export-control fallout

Key points:

There was no jailbreak: Katie Moussouris (Luta Security), the only outside expert to read the report, confirms researchers took code with known CVEs plus planted vulnerabilities and asked models to "fix this code"; Fable provided no uplift over Opus 4.8 or GPT-5.5.
Bloomberg published the Lutnick letter establishing an individually-validated license requirement for any export/transfer of Mythos or Fable worldwide (including deemed exports to foreign employees); WSJ confirms Lutnick told Amodei "that's the point" re: forcing takedown.
Amazon's role: White House likely asked Amazon to test Fable; Amazon found the "jailbreak" and Jassy reported it to Bessent; game-of-telephone between non-technical people misrepresented a harmless defensive capability as a threat.
The takedown also cut off Project Glasswing, UK AISI (Starmer denied a carveout), and Anthropic's internal Mythos access; extending controls to Mythos (not just Fable) was gratuitous.
Zvi maintains Anthropic should still have taken Fable down when ordered (realpolitik precedent), but argues a legal challenge on providing inference might win on statutory (rule 744.22) and constitutional (1A/5A/non-delegation) grounds; WSJ editorial asks "Why Does Trump Hate Anthropic?"

Connections: Zvi Mowshowitz · Anthropic · Amazon · Dario Amodei · Export Controls · Rule Of Law · Cybersecurity