📰 News
Anthropic Mythos 5 ban partially lifted for 100+ US organizations
Two weeks of complete lockdown eased slightly. The Trump administration has authorized Anthropic to restore Mythos 5 access for over 100 specific U.S. government agencies and companies — non-American employees at those orgs included, which is worth noting. Anthropic confirmed: “Mythos 5 is being restored to critical infrastructure defenders; we’re continuing to work on Fable 5.” — TechCrunch
GPT-5.6 Sol, Terra, and Luna debut — limited to partners at government request
OpenAI finally unveiled its next-gen GPT-5.6 lineup. Sol (flagship) brings stronger agentic capabilities in coding, biology, and cybersecurity, with new “max” and “ultra” reasoning modes. Pricing: Sol at $5/$30 per million tokens, Terra at $2.50/$15, Luna at $1/$6. The catch: a government request has kept the launch limited to “trusted partners” for now. OpenAI made its displeasure clear: “We don’t believe this kind of process should become the long-term default.” — TechCrunch
Asian AI startups filling the Mythos vacuum
While Anthropic’s export ban dragged on, Tokyo-based Sakana AI launched “Fugu” (blowfish) and China’s 360 unveiled “Tulongfeng” — both claiming to match Mythos/Fable-level performance. Sakana says the timing was “entirely coincidental,” but its website now prominently features “delivering frontier capability without the risk of export controls.” Fugu also introduces an orchestration model that coordinates multiple AI providers. — TechCrunch
Apple Vision Pro VP Paul Meade heading to OpenAI’s hardware team
Paul Meade, the Apple VP who ran the Vision Pro — and who reportedly also led development of Apple’s upcoming smart glasses — is joining OpenAI’s hardware team. That’s the Jony Ive project building what Sam Altman has described as an AI device more “peaceful and calm” than an iPhone. The departure seems tied to Apple’s internal shuffle following John Ternus’ rise to CEO. — TechCrunch
OpenAI hires Uber India chief as first India Managing Director
Prabhjeet Singh, who ran Uber’s India and South Asia operations, joins OpenAI in September. OpenAI has already been on an India hiring spree and now treats the country as its second-largest market after the US. Anthropic is making the same bet — it opened a Bengaluru office in late 2025. — TechCrunch
CUGA — IBM’s open-source enterprise agent harness, 24 working example apps
IBM Research dropped a detailed HuggingFace post about CUGA (Configurable Generalist Agent). The framework handles the planning loop, tool execution, variable tracking, and multi-agent delegation so you just supply a tool list and a prompt. It topped AppWorld and WebArena benchmarks, and IBM published 24 single-file FastAPI example apps — from a movie recommender to an IBM Cloud architecture advisor. Worth a look if you’re building agent workflows. — HuggingFace
🔥 GitHub Trending
- google-labs-code/design.md — A file format for conveying your design system to coding agents. Think CLAUDE.md but for design: YAML front matter holds design tokens (colors, typography, spacing), markdown prose explains the why behind each choice. Agents read it to produce consistent UI without guessing.
- xbtlin/ai-berkshire — A value investing research framework built on Claude Code. It pits four investment philosophies (Buffett, Munger, Duan Yongping, Li Lu) as parallel agents, surfaces their contradictions explicitly, and forces a concrete buy/pass verdict. Claimed live returns: +69% in 2024, +66% in 2025.
- hugohe3/ppt-master — AI that converts any document into a natively editable .pptx — not slide images — with animations, speaker notes, and optional audio narration. Supports custom templates.
- commaai/openpilot — Open-source robotics OS that upgrades driver assistance on 300+ supported cars. A community perennial that keeps reappearing in trending.
▶️ YouTube
- “The best thing since OpenClaw” — Hermes Tutorial (Matthew Berman) — A full walkthrough of Hermes, an AI agent tool that Berman is clearly excited about. Strong claim in the title; worth a watch if you use coding agents.
- GPT-5.6 officially NOT BANNED (but maybe worse) (Matthew Berman) — Quick take on why “restricted preview” might be more concerning than an outright ban.
💬 Community
- What happened after 2,000 people tried to hack my AI assistant — Fernando Irarrázaval ran a public hacking challenge on hackmyclaw.com: could anyone leak secrets from his Opus 4.6-powered assistant via email? 6,000 attempts, $500 in tokens, one Google account suspension — and zero successful leaks. Simon Willison’s note is the right takeaway: “6,000 failed attempts provides no guarantees that someone with a more sophisticated approach couldn’t get through.”
- Incident Report: CVE-2026-LGTM — A satirical but painfully plausible fictional incident report by Andrew Nesbitt. Two AI code review agents from competing vendors get into a disagreement loop over a single PR bump, rack up 340 comments and $41,255 in inference spend before Finance revokes their API keys. The vendor’s marketing team files a press release about “430% YoY increase in adversarial multi-agent security reasoning.” Stock opens up 6%. Funny until it isn’t.
- AI Learns the “Dark Art” of RF Chip Design — Analog and RF chip design has long been a domain where decades of human intuition couldn’t be automated. IEEE Spectrum reports AI is starting to crack it. (via Lobsters)