StationX
MEMBERSHIP

Sec & AI News — 10 July 2026

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

🟣 GPT-5.6 lands, and it's a real jump

OpenAI shipped GPT-5.6. Three tiers: Sol at the top, Terra and Luna under it. The version number says minor update; using it says otherwise. It coded a full 3D game from one prompt in my testing, and built itself a showoff website from another, roughly Fable-class work for a fraction of the price. The naming, though. Sol, Terra and Luna each come in Ultra, Pro and Light, so that's nine names before you've typed a word.

🟢 ChatGPT eats Codex, Atlas, and everything else

OpenAI folded the Codex app, the ChatGPT app, and the Atlas browser into one thing and called it ChatGPT. Work mode for getting stuff done, Codex mode for shipping code. The browser now lives inside ChatGPT. So do "Sites" — prompt a website into existence and it hosts it, no databases, no CDN, no thinking about any of it. Point it at your Gmail, Drive, Slack, and second brain and it builds a control tower of everything you've been ignoring. Which is helpful right up until you realise it read all of that.

🟤 "Walk away and it builds the whole thing" is a real skill now

Everyone's demoing the same trick this week: hand the model a brief, walk off, come back to a finished platform. I've run this way for over a year, and about 80% of my company now runs on AI infrastructure I built exactly like that. But walking away isn't the skill. The skill is coordinating spiky, half-reliable agents at speed without shipping the vulnerabilities vibe coding quietly sneaks in. I wrote up what Karpathy actually means by agentic engineering, why most people never climb past tier one of four, and the discipline that keeps the whole thing safe — including a real five-agent security review that went request to production in four sentences.

🔵 GPT Live talks over you now

New voice mode. It listens and speaks at the same time, interrupts you, and lets you interrupt it back. Drop your phone on the table between two people who don't share a language and it translates in real time, talking over each speaker as they go. Full-duplex, so the awkward walkie-talkie pauses are gone. It lives under Settings → Voice → model: Live.

⚫ Grok 4.5 shows up better than expected

xAI dropped Grok 4.5. Nobody talks about Grok, and yet here it is on DeepSWE and Terminal Bench, trading blows with Fable Max and GPT-5.5. Install the CLI, run it in your terminal, free for now. API is $2 per million in, $6 out, which is cheap next to Sol's $5/$30. It won't take the crown off anyone this week, but it's a lot closer than Grok's reputation suggests.

🟠 Meta remembers it makes models

Meta shipped Muse Spark 1.1. It lands on par with Opus 4.8 and GPT-5.5 on terminal bench, a step behind the newest releases but a real jump from where Meta was. Going from Llama Scout to this in a year is the actual story, given how thoroughly everyone had written them off. Press pins the API around $1.25 in / $4.25 out.

🔴 Muse Image will render your face without asking

Meta's other launch is the one to watch. Muse Image is free in Meta AI and it connects to Instagram — @-mention anyone with a public account and it generates images of them. Anyone. People you've never met. There's reportedly an off switch buried in settings, except not everyone has it yet. So the consent toggle is shipping after the feature that needs it.

⚫ A model that renders anyone's face is the small version of this

Muse Image generating strangers on demand is the loud version of a much quieter problem. Your face, your name, your movements are becoming a single digital identity, and nobody sat down and designed it that way. The China "social credit score" everyone fears doesn't actually exist the way you were told. What's real is the West assembling something worse by accident, one reasonable convenience at a time. My piece takes apart the myth, then follows the architecture that's actually being built across the US, UK, and Europe right now, close enough that you'll recognise the moment it shows up near you.

🟡 Claude gives you five more days of Fable 5

Fable 5 access was ending July 7. It's now extended through July 12. Then everyone pointed out they'd already burned their weekly limit racing the original deadline, so Anthropic reset the limits too. All of this landing the same week GPT-5.6 and the ChatGPT superapp shipped, which I'm sure is a coincidence.

🟣 Claude Cowork follows you off the laptop

Until now Cowork lived on your machine, so closing the lid stopped the work. Now it runs in the cloud. Start a task at your desk and pick up the finished output from your phone hours later. Convenient — and conveniently timed against the new ChatGPT app doing exactly this, the same week.

🟢 Claude, but make it Spotify Wrapped

New Reflect feature in Claude web and desktop. Most active day, peak hour, conversation count, a breakdown of what you actually spent your time on. It's usage stats with a nicer graph. Most people will open it exactly once.

🔵 Anthropic found the model's inner monologue

Interpretability research from Anthropic on what they call the global workspace, informally J-Space. When Claude answers, you see its chain of thought. This is the other stuff — what the model holds in mind but never says out loud. Anthropic is now tracking those hidden states and finding patterns in them, and framing it as the model's subconscious. The framing happens to line up with a leading theory of human consciousness, which is either profound or a very good way to get a paper read. Your call.

🟢 You have been told to "secure the LLM." Now what?

Anthropic peering into the model's hidden workspace is fascinating, and a reminder that most people deploying these things can't see inside them at all. Every LLM security guide hands you a wall of scary attacks — prompt injection, jailbreaks, data poisoning — then leaves you with no idea what to actually build on Monday morning. So I wrote the one that doesn't. Most of securing an LLM is the security methodology you already know, pointed at a new component that happens to have a few genuinely new failure modes. My guide walks one realistic deployment through seven steps, naming the actual control at each, all mapped to the OWASP LLM Top 10 so nothing slips through.

⚪ Google Photos remixes your memories

Google Photos got Video Remix, powered by Gemini Omni — turn your clips into shareable, stylized videos. Extra flourish and "imagined memory" bolted onto footage you already have. Google AI Plus, Pro, and Ultra only.

🟠 Seedream 5.0 Pro does the things image models usually can't

ByteDance shipped Seedream 5.0 Pro. The pitch is the hard stuff image models usually botch: dense infographics with legible text, interactive editing where you circle and draw to direct it, Photoshop-style layers you can rearrange and re-prompt, plus clean photorealism. Playable inside Leonardo. The layer separation is the bit worth caring about, if it actually holds up outside a demo reel.