StationX
MEMBERSHIP

Sec & AI News — 19 June 2026

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

🟥 What They Don't Tell You About the Model the Government Just Killed

I tested Fable 5 the week it launched — before anyone could. The most capable model ever released, and it refused to do security work, downgrading you to Opus 4.8 the second you mentioned cyber. The unlocked version (Mythos 5 / Project Glasswing) went to about a dozen giant companies. Not you. I called this exact outcome in the review — and here it is, sooner than expected. The lesson: a model that can be switched off overnight by a directive you didn't see coming isn't capability you can build on. The answer is sovereign AI — capability you own and run yourself.

🟧 The Code Review Trick That Catches What One AI Misses

GLM-5.2 writes code. So does everything now. The bottleneck moved to catching what the code gets wrong — and one AI model reviewing code has the same flaw as one human: a single set of blind spots, confidently wrong with no second opinion. The fix is a panel — Claude, GPT and Gemini hunting different surfaces, judged adversarially, every fix verified. Thirty years of finding the bugs that pass the demo and break in production, turned into a workflow.

🔴 The US Government Switched Off Anthropic's Best Model Overnight

First time ever: Washington forced a publicly released commercial model off the market. On Friday 12 June, Anthropic had to kill Claude Fable 5 and Mythos 5 — for everyone on Earth. The order was narrow on paper: block all foreign nationals, worldwide, including Anthropic's own staff. Impossible to enforce surgically, so Anthropic pulled the plug for all of it. The irony writes itself. Dario Amodei spent months calling this class of model a national-security risk, then published an essay arguing the government should be able to block or reverse a release. The government read the essay and did exactly that. The reported trigger: a jailbreak flagged by Amazon researchers — Amazon being one of Anthropic's biggest investors. Anthropic called the flaw trivial. The government called that the problem. 76 cybersecurity veterans signed a letter demanding the models back, arguing you don't disarm the defenders mid-fight. Opus 4.8 stayed live throughout.

🟢 An Open-Weight Model Just Crashed the Frontier Party

Z.ai shipped GLM-5.2. Open weights, MIT license, do what you want. 753B parameters, ~40B active, 1M-token context. Built for long-horizon agentic coding. The numbers: ~62% on SWE-bench Pro against Opus 4.8's ~69%. Not the new king. But it's the best open-weights model on the board — and it costs roughly a sixth of the closed frontier per token. The gap between "open" and "frontier" used to be a chasm. Now it's a rounding error with a price tag attached.

🔵 Midjourney's Next Product Is a Body Scanner

The image-generation company built a medical imaging division. "Midjourney Medical" — a radiation-free full-body ultrasonic scanner that builds a 3D body map in about 60 seconds, on Butterfly Network's ultrasound-on-chip tech. The plan isn't hospitals. It's the "Midjourney Spa" — first one opening in San Francisco late 2027, scanners next to the saunas and cold plunges. Get scanned, then go sit in the hot tub. CEO David Holz: sensor company, then image generation, now this. Hank Green called out the marketing for comparing it to MRI and CT — ultrasound is great at soft tissue, useless through bone and air. Different tools, different jobs. Cool machine. Just not the MRI killer the deck says it is.

🟠 Codex Will Now Watch You Work, Then Do It For You

OpenAI shipped "Record & Replay" in Codex. Record yourself doing a task once — clicks, typing, up to 30 minutes — and Codex turns it into a reusable skill. Next time, drop in the files and tell it to run the skill. The demo: uploading a YouTube video. Watch once, repeat forever. macOS only, needs Computer Use, blocked in the EU/UK. Macro recorder with a brain. We've come full circle to "automate the boring stuff," except now it learns the boring stuff by watching.

🟣 Claude Design Lets You Edit on the Canvas

Anthropic updated Claude Design. Generate a design, then edit it directly — per-element WYSIWYG, sliders, inline comments. Import your design system from GitHub repos, design files, or raw uploads, and it builds against your components and self-corrects before you see the output. Round-trips to Claude Code via `/design-sync`, plus connectors out to Canva, Replit, Vercel, Wix and friends. On the paid plans. The "build with your components and check itself" part is the actual story — everything else is table stakes now.

🟡 Perplexity Gave Its Agent a Memory That Improves Itself

Perplexity launched "Brain" for its agent product, Computer. It builds a context graph of the work it does, then reviews that graph overnight and teaches itself to do the job better. Learns from your corrections. Remembers which sources were dead ends. The pitch: fewer turns, fewer model calls, better output the more you use it. Reported gains — +25% correctness, -13% cost. Same idea everyone's chasing right now: an agent that gets less stupid the longer you keep it around. Max and Enterprise plans only.

🟤 Adobe Put an AI Assistant Inside Premiere, Illustrator and InDesign

Adobe's Firefly AI Assistant is in public beta across Premiere Pro, Illustrator and InDesign (plus Photoshop and Frame.io). Talk to it instead of clicking. "Sort and bin the drone footage." "Generate this image." "Version these files from the spreadsheet." The editing stays where it always was — you just stop doing it by hand. For anyone who's ever rotoscoped a clip at 2am, the appeal is obvious.

⚫ Google Shipped "Ask Ad Manager"

Gemini-powered agent inside Google Ad Manager, aimed at publishers. Ask it how your inventory's performing, troubleshoot delivery, generate reports, navigate the platform in plain English. Same energy as "Ask Studio" in YouTube — point the chatbot at your numbers, get told what's broken and what to fix. The natural endpoint for every dashboard: stop reading it, just ask it.

🔘 Half of America Uses AI. Most of Them Don't Trust It.

Pew surveyed 5,119 US adults. 49% now use AI chatbots — up from 33% in 2024. ChatGPT leads at 44%, still the Kleenex of AI. But the same people are skeptical: more expect a negative societal impact than a positive one, and the heaviest users — under-30s — are also the most worried. Both things true at once. Because "AI" isn't one thing. ChatGPT saves you an hour; the AI slop on your feed ruins ten minutes. You can like one and hate the other. The survey just can't tell the difference.

⚪ Someone Built a Self-Driving Toilet

Chinese firm Yueban showed the Xiaoban — an autonomous toilet robot. It drives to your bedside on LiDAR, has a warm-water bidet, cleans and disinfects itself, then drives back, grinds and flushes the waste into your plumbing, and docks to recharge. Built for the elderly and mobility-impaired, which is a genuinely good use of robotics. Around $4,000. The dignity case is real. The "lazy rich person summons toilet to the couch" case is also, regrettably, real.