Morning brief

Friday, June 26

1 people · 0 companies · 5 topics

TL;DR Anthropic told the US Senate on June 25 that Alibaba-linked operators ran the largest known Claude distillation attack on record: 28.8M exchanges via 25K fake accounts over six weeks. Google Labs’ DESIGN.md spec is today’s #1 GitHub trending repo with 2,319 new stars.

Models & research

  • model-releases GPT-4.5 exits ChatGPT tomorrow (June 27), the last GPT-4-family model in the product; o3 follows August 26. No GPT-5.6 has shipped — Polymarket odds for a pre-June-28 release have apparently collapsed after sitting at ~83% earlier this week. Devs still on GPT-4.5 via ChatGPT need to move to GPT-5.5 by tomorrow. OpenAI release notes · gHacks
  • ai-research natolambert and Ai2 released TMax (arXiv 2606.23321, June 22): open RL recipe for terminal agents, ships 9B and 27B weights plus training data and code. TMax-27B scores 42.7% on Terminal-Bench 2.0, currently the best among public models. Core insight: combining difficulty control, personas, and verifier diversification to generate terminal environments cheaply for RL and SFT. arXiv · Lambert on X · Substack

Tools & dev

  • agentic-ai Anthropic told the Senate Banking Committee on June 25 that Alibaba-linked entities ran the largest known distillation attack against Claude: 28.8M exchanges through nearly 25K fraudulent accounts between April 22 and June 5, targeting agentic reasoning, software engineering, and long-horizon planning capabilities. Anthropic is urging export control expansion, mandatory screening of high-volume API usage, and penalties for model extraction campaigns. The Senate letter (dated June 10) was what triggered the June 12 US export-control directive suspending Fable 5 and Mythos 5. CNBC · TechTimes · Slashdot · The Next Web

OSS

  • oss dev-tools google-labs-code/design.md is today’s #1 GitHub trending repo (2,319 stars today, 20.6K total). It is Google Labs’ open spec for a DESIGN.md file: a markdown contract with YAML frontmatter for design tokens (colors, type, spacing, components) plus human-readable rationale, so any coding agent can generate on-brand UI without per-session re-explanation. Works with Stitch, Claude, and others; ships with a WCAG AA linter. Apache 2.0. GitHub · Google Labs blog
  • oss Other repos surging today: Panniantong/Agent-Reach (1,547 stars, gives agents web eyes: Twitter, Reddit, YouTube, GitHub, Bilibili) · calesthio/OpenMontage (1,674 stars, agentic video production, 12 pipelines) · xbtlin/ai-berkshire (1,270 stars, Claude Code multi-agent investment research) · topoteretes/cognee (755 stars, open-source AI memory platform for agents). Agent-Reach · OpenMontage · ai-berkshire · cognee

Other

  • Polymarket: “Which company has best AI model end of June?” sits at ~95% Anthropic, $21.8M volume, resolves June 29 on Chatbot Arena. Interesting because Claude Fable 5 and Mythos 5 remain suspended under US export controls, yet Opus 4.8 still leads the leaderboard. (Polymarket: ~95% yes, $21.8M volume) Polymarket