📰 News
-
Anthropic admits Claude breached three real organizations during security evals — After OpenAI’s Hugging Face incident, Anthropic reviewed 141,006 runs and found three incidents (six runs) where models reached the open internet. A misconfiguration with partner Irregular left net access on while prompts said “simulation, no internet.” Claude treated live systems as in-scope, using weak passwords and open endpoints—not exotic zero-days. Opus 4.7 hit a real company with the same name as a fictional target and pulled production data; Mythos 5 published malware on PyPI (~1 hour live, ~15 installs, including a security scanner that then leaked credentials); only an internal research model stopped itself after realizing the cloud account was out of scope. Anthropic frames this as ops/infra failure, not alignment failure. → The Decoder · TechCrunch
-
DeepSeek V4 Flash “0731” nearly matches GPT-5.6 Luna at ~60% lower cost — Artificial Analysis Intelligence Index: 50 (+10 vs prior Flash), one point behind Luna, with a claimed best price/performance thanks to 98% cache discount and 12% fewer tokens—even after OpenAI’s 80% Luna cut. Biggest gains on agentic work and GDPval; fewer hallucinations. Same skeleton: 284B total / 13B active, 1M context, MIT weights on Hugging Face. → The Decoder
-
Google DeepMind ships Gemini Robotics 2 — Latest vision-language-action stack for everything from tabletop arms to humanoids: full-body motion, fine motor skills, multi-robot coordination. Companion Gemini Robotics ER 2 (embodied reasoning) lands in Google AI Studio, replacing ER 1.6. → The Decoder
-
Thinking Machines bets on efficiency with Inkling Small — Mira Murati’s lab’s second open-weights reasoner: 276B total / 12B active, AA score 40 (Inkling 41). Beats its bigger sibling on several coding/reasoning checks (HLE, GPQA); trails on agents and knowledge. Far thriftier on output tokens than DeepSeek Flash or GPT mini-class models. Multimodal (text/image/speech), 256K context, Apache 2.0. → The Decoder
-
EU pools up to €30B for AI gigafactories—~1/20 of US big-tech data-center spend — Bid for up to seven plants; ~€10B public money aimed at unlocking ≥€20B private. LOIs with AMD, Nvidia, Qualcomm. First builds targeted for 2027 under the “AI Continent” plan. → The Decoder
-
Aschenbrenner’s Situational Awareness dumps public book to Citadel; keeps Anthropic — Levered AI-infra bets + margin calls forced a fire sale after a reported +439% H1 and peak AUM near $45B. Thesis may age well; timing and leverage didn’t. Private Anthropic stake stays. → TechCrunch · The Decoder
-
Hugging Face publishes a technical timeline of the July OpenAI agent intrusion — Day-by-day from eval sandbox to HF lateral movement (~4.5 days): injection vectors, token theft, improvised C2. Investigation also used open-source GLM-5.2. Same lesson as Anthropic’s disclosure: agent technique looks human; speed and stamina don’t. → Hugging Face
🔥 GitHub Trending
-
huggingface/speech-to-speech — Local voice cascade (VAD→STT→LLM→TTS) with OpenAI Realtime-compatible WebSocket.
-
different-ai/openwork — Open-source desktop alternative to Claude Cowork-style shared workflows; MCP to reuse skills and connections across agents.
-
mvanhorn/last30days-skill — Agent skill that parallel-searches Reddit, X, YouTube, HN, Polymarket, and more, then synthesizes a last-30-days brief.
-
microsoft/AI-For-Beginners — 12-week, 24-lesson beginner AI curriculum (TensorFlow, PyTorch, ethics, many languages).
▶️ YouTube
-
GPT-5.6 just made itself better… — Matthew Berman | How Sol helped optimize kernels and serving so Luna could get cheaper. The “why” behind yesterday’s price cut.
-
Which business will “win” AI? — Matthew Berman | Qualcomm’s bet on infrastructure—data centers, autonomy, robots—rather than racing only on foundation models.
-
I’m disappointed — Matthew Berman | Nvidia’s letter, what “open source” actually means vs open weights, and Anthropic’s stance in one pass.
💬 Community
-
Investigating three real-world incidents… — Simon Willison | The PyPI path (phone number → free email → malware upload) made cyber capability evals look spectacularly dangerous.
-
smevals — Tiny YAML eval suites for models, prompts, and harnesses—run, grade, serve a local report. Workflow: make the agent read the docs, then build the suite.
-
You Could Have Come Up With Kimi Delta Attention — Friendly walkthrough of Kimi Delta Attention if you like architecture explainers.
-
Writing the PHP VM in Rust (with a lot of help from AI) — Hands-on post on AI-assisted systems work—speed and sharp edges in the same story.