AI/ML Digest

Morning Edition Sat Jun 13

The U.S. government issued an export control directive on June 13 ordering Anthropic to disable Claude Fable 5 and Claude Mythos 5 for all customers worldwide, citing a reported jailbreak as a national security risk. Anthropic complied at 5:21 PM ET but stated it believes the directive is based on a misunderstanding, as the demonstrated jailbreak appeared narrow and non-universal.

newsU.S. government orders Anthropic to shut down Fable 5 and Mythos 5

Export control directive disables Anthropic's two most capable models globally, citing a jailbreak national security risk. — TechCrunch AI

newsAnthropic publishes full statement on Fable 5 and Mythos 5 suspension

Anthropic's statement details the directive, its jailbreak assessment, and why it disputes the government's evidence. — Simon Willison

newsFable 5 and Mythos 5 access revoked for all customers after export order

AINews roundup covers the Anthropic shutdown and broader open-source AI community reaction to the export control action. — Latent Space

modelNVIDIA GB300 NVL72 delivers 20x agents-per-megawatt over Hopper on AgentPerf

Artificial Analysis's new agentic benchmark shows Blackwell Ultra NVL72 leading on chained LLM and tool-call workloads. — NVIDIA Blog

newsGoogle sues Chinese cybercrime network Outsider Enterprise over AI-powered scams

Outsider Enterprise ran 9,000 fake sites and sent 2.5M scam texts in two weeks; FBI seized domains in coordination. — TechCrunch AI

techniqueMicrosoft Project Ire autonomously classifies novel LOTUSLITE malware variant

LLM-driven agent produced a full behavioral report on an undetected DLL backdoor with no human input or prior signatures. — Microsoft Research

toolaisuite adds OpenCoworker desktop agent with multi-provider support

Lightweight Python library now ships a desktop agent harness supporting OpenAI, Anthropic, Google, and local Ollama models. — GitHub Trending

toolLlaboratory: open-source testbed for LLM tool-call behavior

Self-hosted Docker tool lets you create fake tools, run versioned test plans, and inspect model tool-call traces. — HN Show HN

toolPaca: open-source Scrum board treating AI agents as first-class teammates

Self-hosted project management platform lets AI agents join sprints, pick tasks, and write specs alongside humans. — Hacker News

toolcc-pool load-balances Claude Code sessions across multiple subscriptions on macOS

Go tool picks the least-used Claude account before each session starts, keeping secrets in the macOS Keychain. — HN Show HN