The Once And Future Fable #4
ai-policyexport-controlsanthropiccybersecurityai-governance
Abstraction: Fable takedown aftermath, cyber-defense, restoration odds
Key points:
- Fourth update on the US government forcing Anthropic to take down Claude Fable 5 / Mythos 5; Polymarket gives 60% chance of restoration by July 1, 88% by July 31, with Claude Code v2.1.190 strings hinting Fable may return permanently within subscriptions with weekly usage quotas.
- Debunks the viral "Mythos broke into almost all NSA classified systems in hours" claim: it was an authorized NSA red team with physical access to air-gapped systems using Mythos alongside Cobalt Strike; Shashank Joshi (who wrote the original Economist line quoting Sen. Warner) says Warner misunderstood Gen. Rudd, and red teams lost Mythos access (was under Project Glasswing).
- Argues restricting deployment does not slow development and may speed it up by freeing resources; Andrew Curran reports a more capable Mythos (5.1 or 6) already emerged from training.
- Extended debate with Arthur Tellis/Teortaxes on offense-vs-defense in superhuman hacking: Zvi rejects "cybersecurity risk is a small tax" framing, arguing cybersecurity value equals all of IT value; defends data-retention + KYC + classifier observability as defense-in-depth.
- Coins "Anthropic Derangement Syndrome" (ADS) re: Ben Thompson attributing maximally negative motives to Anthropic; endorses davidad's point that deployment-restriction agreements are easy to monitor/enforce unlike development halts.
Connections: Zvi Mowshowitz · Anthropic · Claude · Export Controls · Cybersecurity · AI Policy
Source: https://thezvi.substack.com/p/the-once-and-future-fable-4