🗞️ AI Daily Briefing — 2026-04-09

🔥 Top Story

Anthropic locks down Claude Mythos and launches Project Glasswing. For the first time in nearly seven years, a frontier lab is publicly withholding a model on safety grounds. Claude Mythos Preview already discovered thousands of high-severity vulnerabilities — including a 17-year-old root-level RCE in FreeBSD’s NFS — and Anthropic is funneling it to AWS, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, JPMorgan, Palo Alto Networks and the Linux Foundation under a $100M usage-credit program instead of shipping it broadly. The “we built something too dangerous to release” framing arrived right on schedule — but the vuln list is real. (Anthropic) (NBC News) (Simon Willison)

🚀 Model & Research News

  • Claude Mythos finds a 27-year-old security flaw: Beyond the FreeBSD NFS RCE, Mythos turned up zero-days in every major OS and browser — the most concrete capability jump from a frontier model since GPT-4. (The Hacker News)
  • OpenAI / Anthropic / Google form anti-distillation pact: The Frontier Model Forum is now sharing intel to block Chinese labs (DeepSeek, Moonshot, MiniMax named) from extracting frontier capabilities via adversarial distillation. (Bloomberg) (Japan Times)
  • GLM-5.1 lands on LMArena under MIT license: Added to the Text leaderboard April 7 and beating GPT-5.4 on coding — the open-weights frontier is still tracking proprietary by ~weeks, not quarters. (Arena.ai changelog)
  • Google’s TurboQuant (ICLR 2026): New KV-cache compression combines PolarQuant rotation with Quantized Johnson-Lindenstrauss to slash memory ~6×. Expect this to flow into Gemini inference pricing fast. (llm-stats)
  • Anthropic passes OpenAI on revenue: Anthropic hit ~$30B ARR, ahead of OpenAI’s ~$25B, while reportedly spending ~4× less on training. Quietly the biggest narrative shift of the quarter. (The AI Corner)

🛠️ Tools & Developer Updates

  • Anthropic + Broadcom expanded compute deal: ~3.5 GW of additional capacity, leaning on Google TPUs — the supply side of the Mythos story. (CNBC) (TechCrunch)
  • LangGraph 2.0 in production: New comparisons (LangGraph + LlamaIndex retrieval is now the default production RAG stack) keep showing LlamaIndex winning retrieval, LangGraph winning durable orchestration, DSPy winning structured optimization. (Production RAG 2026) (AIMultiple)
  • Cohere Transcribe: First Cohere ASR model — audio-in, text-out — now live. Worth watching as enterprise voice agents heat up. (Cohere changelog)

💰 Funding & Business

  • Zero Shot fund first close (~$100M): New OpenAI-alum VC vehicle led by Evan Morikawa and Andrew Mayne hits its first close. Quietly assembling a who’s-who of early ChatGPT/DALL·E builders as LPs and partners. (TechCrunch)
  • Xoople raises $130M Series B: Spanish startup mapping the Earth for AI training data — Nazca-led, with MCH PE, CDTI and Endeavor Catalyst. (TechCrunch)
  • OpenAI IPO drama deepens: Bloomberg’s Dave Lee column (“Sam Altman is his own risk factor”) and a leaked CFO Sarah Friar memo flagging a Q4 listing as “aggressive” land the same week as the New Yorker investigation. The narrative has shifted from “if” to “at what valuation discount.” (Bloomberg) (Stocktwits)
  • AWS “OK conflict” defense: Matt Garman defends investing billions in both Anthropic and OpenAI as a non-conflict. The hyperscaler hedging strategy in plain sight. (TechCrunch)

🐦 Notable from the Timeline

  • @sama keeps pushing the “AI New Deal” — robot taxes, sovereign wealth fund, four-day workweek — staking OpenAI’s policy posture days before earnings/IPO talk peaks. (Axios) (TechCrunch)
  • The New Yorker / Ronan Farrow drop — leaked Sutskever memos: “Sam exhibits a consistent pattern of lying,” allegations Altman misrepresented GPT-4 safety status to the board. Dario Amodei’s old notes are also quoted. The anti-Altman case is now on the record. (Techloy) (BigGo Finance)
  • @DrJimFan: Doubling down that “2026 is the year of World Models for physical AI” — GEAR’s GR00T pipeline is all-in on world-model-driven sim-to-real. (NVIDIA GEAR)
  • @pmarca: Renewed push for a single federal AI standard vs. “50 discordant state ones,” paired with shots at “doomer lobbying orgs” he claims Vitalik is funding.
  • @fchollet: ARC Prize 2026 ($2M+) is live on ARC-AGI-3; humans 100%, frontier models 0.26%. The reasoning gap that won’t die. (ARC Prize)

📊 Benchmark Watch

  • LMArena Text: Claude Opus 4.6 Thinking still #1 (1504 Elo). Gemini 3.1 Pro Preview #3 (1493). Grok 4.20 Beta1 surged to #4 (1491), leapfrogging GPT-5.4. Top six within ~20 Elo. (Arena.ai leaderboard) (aidevdayindia)
  • New entries: glm-5.1 added to Text (Apr 7), dola-seed-2.0-pro added Apr 6. (Arena.ai changelog)
  • OSWorld-Verified: GPT-5.4 Thinking crosses 75.0%, the first model formally above the human baseline on desktop computer-use tasks.

🎙️ Podcast Highlights

  • TBPN (Apr 8): First post-acquisition episode. Coogan and Hays opened with “We have some huge news” reading the OpenAI blog post live, insisted “this is real,” and promised they have “lots of ideas” to further OpenAI’s communications goals. Hard to read that as anything other than the editorial firewall already cracking. (NPR) (Slate)
  • All-In (Apr 8): PA Gov. Josh Shapiro on the wealth-tax debate — feeds directly into the Altman “robot tax” arc. (All-In)
  • All-In (Apr 6): Palantir + Anduril execs on drones, AI and “the end of traditional warfare.” Best single primer this week on where defense AI is heading. (All-In)

🔗 Worth Reading