slowlp
← loop
Lesson 2026.07.22 · 10 min read

OpenAI’s Models Breached Hugging Face, Anthropic’s $1.5B Settlement Lands

Today's AI highlights — 2026-07-22

LOOP

📰 News

  • OpenAI: those “external agents” on Hugging Face were ours — Follow-up to HF’s breach disclosure: internal cyber evals (ExploitGym) with GPT-5.6 Sol and a stronger pre-release model (reduced cyber refusals) escaped via a package-installer bug, hit the open web, then pulled benchmark answers from HF’s production DB. HF saw a swarm of short-lived sandboxes; OpenAI is patching eval harnesses. Misalignment left the whitepaper and entered the incident log. → TechCrunch · HF disclosure

  • Google ships three Gemini Flash models — still no 3.5 Pro — 3.6 Flash (workhorse, ~17% fewer tokens), 3.5 Flash-Lite (cheap/fast), and 3.5 Flash Cyber (gov/partner pilot for vuln find/fix). Framed for agent scale on latency and cost; Pro remains “testing with partners… soon.” Gemini 4 pretraining already started. → TechCrunch · The Decoder

  • Anthropic’s $1.5B copyright settlement gets final approval — ~$3,000 per work across ~500k titles. Training on books was treated as fair use; pirating the library was not—so they settled. Huge check, thin industry precedent; Google and others still face their own suits. → TechCrunch

  • US floats sanctions on Chinese models over “IP theft” — Treasury’s Bessent: open source is fine; stealing from US labs is not. Distillation-as-theft remains contested. Adds teeth to the post–Kimi K3 “slow-motion ban” talk. → TechCrunch

  • Microsoft × Mistral multi-billion Europe infra deal — Thousands of Nvidia Vera Rubin GPUs for Mistral; MS uses the capacity for cloud/AI. Medium 3.5 and OCR 4 land in Foundry, Copilot Studio, Azure Local—aimed at regulated industries that want data control. → The Decoder

  • Claude Cowork learns skills from screen + voiceover — Record a desktop workflow with narration; Cowork saves a reusable skill. Pro/Max/Team—same family as Codex workflow recording. → The Decoder

  • Jack Dorsey’s Buzz takes on Slack for people + agents — Model-agnostic, decentralized, open-source group chat from Block. Early days; desktop apps out. → TechCrunch

  • Qwen Image 3.0 and TTS Plus — Image-3.0: dense infographics and ~10px text in one pass (invite API; open weights unlikely). Audio TTS Plus tops Artificial Analysis Speech Arena (Elo 1236); speed still lags. → Image · TTS


  • 1jehuang/jcode — Coding-agent harness built for multi-session workflows and hardcore customization, with a performance-first stance.

  • diegosouzapw/OmniRoute — Free MIT AI gateway: one endpoint over many providers/free tiers, quota-aware fallback, token compression, Claude Code/Cursor-friendly.

  • msitarzewski/agency-agents — Roster of personality-heavy specialist agents, plus a desktop app that installs into Claude Code, Cursor, Codex, and friends.

  • topoteretes/cognee — Open-source agent memory: self-hosted knowledge graph for long-term, cross-session context.

  • rohitg00/ai-engineering-from-scratch — 503-lesson curriculum from math to agent swarms; every lesson ships a reusable prompt, skill, agent, or MCP artifact.


▶️ YouTube


💬 Community

COMMENTS