StationX
MEMBERSHIP

Sec & AI News — 12 June 2026

8 min readBy Nathan House
Get every new Sec & AI News issue
Straight to your inbox. No spam.

🛡️ Our Take: Fable 5 Is a Security Nightmare

Anthropic released Claude Fable 5 on June 9 — its most capable model ever, in a new "Mythos-class" tier above Opus. A 1-million-token context window, runs autonomously for hours, and scores at roughly human senior-engineer level on coding benchmarks. It also finds security holes better than anything ever made — a 16-year-old bug in FFmpeg, a 27-year-old flaw in OpenBSD. Then they banned it from doing security. Touch anything security-shaped — code review, vulnerability research, even researching the model itself — and it quietly swaps you to Opus 4.8. While charging double. The unlocked version, Mythos 5, exists; you can't have it. Unless you're AWS, Apple, Google, Microsoft, or JPMorgan. The gate is a sticky plaster — analysts give it 9-12 months before this capability is everywhere. We've written up the full breakdown: what actually gets flagged, the mid-response double-billing sting, the settings toggle that makes the downgrades visible, and what security professionals should do today.

🟣 Anthropic Ships Fable 5. Apologizes Within Two Hours.

Anthropic launched Claude Fable 5 and Mythos 5 on June 9 — a whole new "Mythos" tier above Opus. Fable 5 is state-of-the-art on most benchmarks, runs autonomously longer than any previous Claude, and costs $10/$50 per million tokens. Double Opus 4.8. You'll pay it anyway. The actual Mythos 5 — same model, fewer safeguards — is reserved for a small club of cyber defenders and infrastructure providers. Everyone else gets the supervised version.

🚧 The Guardrail Problem

The catch: Fable 5 silently rerouted frontier-LLM-development questions to a dumber model. No notice. It also refused blood-work analysis and cancer questions as "unsafe." The Hugging Face CEO called it concentration of power. Jeremy Howard called it the opposite of the safe path. Anthropic reversed course in under two hours — the safeguards stay, but now they'll tell you when you've hit one. Transparency, the sequel.

🍎 Siri Finally Gets a Brain. Google Built Half of It.

WWDC 2026 was wall-to-wall AI. Apple's next-gen foundation models are co-built with Google Gemini — on-device or private cloud, with Google getting none of your data, allegedly. New Siri AI reads your photos, calendar, and messages, gets a dedicated app, lives in Spotlight on Mac, and operates your phone by voice. There's a Siri camera mode, AI-built Shortcuts from plain English, and generative photo reframing in the iOS 27 beta. EU users: not at launch. Regulation has a price, and it's you waiting.

📓 NotebookLM Gets Its Own Computer

NotebookLM now runs on Gemini 3.5, and every notebook ships with a secure cloud computer that writes and executes code. Over 100 curated skills run behind the scenes. Output formats now include PDF, DOCX, XLSX, PPTX, SVG, and editable everything. It went from research notebook to research intern.

🌐 Gemini 3.5 Live Translate: Language Barriers, Cancelled

Near-real-time speech translation. Watch a video in Hindi, hear it in English. It's coming to Google Meet and the Translate app on Android and iOS, and it's live in AI Studio right now. Multilingual meetings just became monolingual meetings with extra steps removed.

⚡ DiffusionGemma: Text Generation, Stable Diffusion Style

Google's new open model generates text the way image models generate pixels — whole 256-token chunks at once instead of one word at a time. Intelligence is roughly Gemma 4 12B. Speed is nothing else's. The point: local inference that actually uses your GPU instead of letting it idle between keystrokes. On-device AI keeps getting more serious.

📜 Dueling Manifestos: Dario and Sam Explain the Future

Both lab CEOs published vision documents this week. Dario Amodei's "Policy on the AI Exponential" wants an FAA-style review body for frontier models, UBI funded by higher AI taxes, a faster biomedical pipeline, an off-switch, and an international coalition. Sam Altman and Jakub Pachocki's "Built to Benefit Everyone" declares OpenAI's third phase: AI doing a significant fraction of their research by March 2028 and a personal AGI for every human. Both promise broadly shared prosperity. Both run companies racing to make it true first.

✉️ ChatGPT Sends Email Now

Draft in chat, plug in an address, hit send. No Gmail tab required. Convenient. Also one more app with send-as-you privileges. What could go wrong.

📈 OpenAI Files for IPO

Confidential S-1, submitted. No date, no price, no numbers — that's the "confidential" part. The intention is the news: the most-watched private company in tech is heading for the public markets, probably next year.

🚀 SpaceX Raises $75 Billion. Largest IPO Ever.

555 million shares at $135. A $1.77 trillion valuation before the first public trade. Trading is now live, so that number belongs to the market now. Records are made to be broken; this one just broke all of them at once.

🤖 Coinbase Hands Your Wallet to Your Agent

Coinbase opened its platform to AI agents — they can now trade and pay on your behalf via MCP. If you trust your agent with your money, the infrastructure is ready. The infrastructure being ready and this being a good idea are separate questions.

🔮 Midjourney Is Making… Hardware?

Founder David Holz is sending out invites for Midjourney's "first hardware launch." No product, no date, no details. An image-generation company building a physical gadget. Speculation is free; whatever it is won't be.