14 Jun 2026

Digest: US export controls pull Anthropic's Fable 5/Mythos 5, CAISI evaluates DeepSeek V4

An unprecedented US government order forcing a frontier lab to disable its most capable models dominates the day, alongside fresh independent evaluations and new open-weight releases.

1. US government directive to suspend access to Fable 5 and Mythos 5

   Citing national-security export-control authorities, the US government issued a directive requiring Anthropic to suspend all access to its Fable 5 and Mythos 5 models for any foreign national—including foreign-national Anthropic employees—forcing the company to abruptly disable the models for all customers while other models stay available. The episode marks an extraordinary case of the government using export controls to pull a deployed frontier model from the market.

   Capybasilisk via LessWrong
2. Claude Fable 5 and Mythos 5: The System Card

   Anthropic's newly released Claude Fable 5 and Mythos 5 system card, analysed by an independent commentator, describes Fable 5 as the strongest publicly available model and notes a step-change in usefulness over prior models. The capability jump and accompanying safeguards detail are central context for the government's subsequent move against the same models.

   thezvi.wordpress.com
3. Washington made a frontier AI model disappear

   Reporting frames the Fable/Mythos shutdown as the de facto arrival of an AI licensing regime imposed through export controls rather than a deliberate policy framework. It raises the question of how governments will gate frontier-model access and what precedent a sudden national-security takedown sets for future deployments.

   Shakeel Hashim via Transformer News
4. CAISI Evaluation of DeepSeek V4 Pro

   The US government's AI evaluation body (CAISI) published an assessment of DeepSeek V4 Pro, continuing its program of independent third-party evaluations of Chinese frontier models. Such government evaluations are a key input for dangerous-capability and national-security risk assessment of widely used foreign open models.

   NIST (AI news)
5. Moonshot AI releases Kimi K2.7-Code open-weight coding model on Hugging Face

   Moonshot AI released open weights for Kimi K2.7-Code, a trillion-parameter mixture-of-experts agentic coding model (32B active) under a permissive license, reporting double-digit gains on coding benchmarks over its predecessor while using fewer reasoning tokens. It marks continued open-weight frontier-tier agentic-coding progress from a major Chinese lab.

6. Investing in multi-agent AI safety research

   Google DeepMind and partners announced a $10M funding call for multi-agent AI safety research, directing resources toward risks that emerge when multiple AI agents interact. The move signals growing institutional investment in a safety area that current single-model evaluations largely do not cover.

   Google DeepMind