OpenAI paces Astra over cyber risk, Anthropic hits $65B ARR

📰 News

OpenAI is pacing Astra over cyber risk — RL paused two weeks after the Hugging Face leak; the biggest frontier RL run is still on hold. Alerts within 30 minutes, monitoring at ~20% of inference. They say Astra is close to critical cyberattack skills. → The Decoder
Anthropic’s annualized revenue hits $65B — $47B in May, $9B at year-end 2025. Investors see $100–120B by year-end, with a confidential IPO aimed at $2T+. OpenAI is at $40B. → TechCrunch
ChatGPT for Teens ships — Ages 13–17. Tighter rules on suicide, self-harm, and sexual content; no pretending to have feelings. Homework gets questions, not answers. Age is inferred from 2,000+ signals. → TechCrunch
Dario Amodei: open models just move power to whoever owns the chips — AI centralizes because of scaling, he says. Baker, LeCun, and Sacks call it fear-driven regulatory capture. → The Decoder
A new leaderboard for agent search APIs — Same GPT-5.6 Luna agent; only the search backend changes. Parallel, Exa, Firecrawl land at 75, 74, 73. No search: 33. Better search cuts tokens 40%+. → The Decoder
Agent memory is a dose, not a switch — Weaker models want curated retrieval (gpt-oss-120b +16.1pp). Strong ones want the full guideline set. GLM-5 sat at zero. → Hugging Face

usestrix/strix — Autonomous AI pentesters. They validate real exploits and drop into GitHub Actions.
akitaonrails/ai-memory — Long-term memory for coding agents. Quit Claude Code, open Codex in the same folder, keep the context.
mukul975/Anthropic-Cybersecurity-Skills — 817 agent skills across 29 security domains. Community project, not Anthropic. Mapped to ATT&CK and NIST.
AlexsJones/llmfit — One command to see which models actually fit your RAM and GPU. Ollama, llama.cpp, MLX.
jundot/omlx — Local LLM server for Apple Silicon. Continuous batching and SSD KV cache, from the menu bar.

Qwen 3.8 27B scores 52 on the AA Intelligence Index — Simon. Tied with GPT-5.6 Luna. One point behind GLM-5.2 (753B) and DeepSeek V4 Pro (1.7T). It’s 27B.
The Limits of AI (1985) — A 40-year-old lecture is back on the feed. Reads differently next to the Astra debate.