Sec & AI News — 3 July 2026
🔌 I Asked My AI to Design Its Own Body
I went the other way from everyone chasing bigger models this week — I asked my AI to design the hardware itself. One morning, one prompt, and under an hour later the manufacture-ready factory files for a real circuit board were sitting in a folder on my screen — printable for about the price of a coffee. Designed almost entirely by AI. The frontier everyone's watching is the models; the quieter one is AI crossing out of the screen and into physical objects you can hold.
- AI PCB Design: My AI Tried to Make Itself Physical (2026) — by Nathan House
🔒 GPT-5.6 Ships — the AI You're Not Allowed to Use
OpenAI shipped GPT-5.6 in three tiers, celestial theme: Sol (flagship), Terra (mid), Luna (fast and cheap). Sol Ultra — a reasoning mode on top — scored 91.9% on Terminal-Bench 2.1, beating Fable, at half the price ($5 in / $30 out per million tokens). It's OpenAI's most capable model yet for coding and cybersecurity — one that finds real flaws in the code behind Chrome and Firefox. And you can't have it. Same playbook that killed Fable 5 last month: at the US government's request, it went to a tiny group of vetted partners via API and Codex, released customer-by-customer, broad access "in the coming weeks." Faster, cheaper, better, rationed by approval list.
- OpenAI — Previewing GPT-5.6 Sol
- VentureBeat — OpenAI unveils GPT-5.6 Sol, Terra and Luna
- The full story of what it does and why it's locked away: GPT-5.6 Sol: The Hacking AI You're Not Allowed to Use (2026) — by Nathan House
⚖️ The Uncomfortable Truth About Who Controls AI Now
In two weeks this June, the US government did something it had never done openly: it walked into Anthropic and OpenAI and told them what they could ship. No law. It reached for export controls — the machinery built for weapons — and treated a frontier model like a munition. Anthropic pulled two models. OpenAI was told to hand its next one out a government-approved customer at a time. That's the real story under every headline this week. Not which model is best — who holds the switch.
- The Uncomfortable Truth About Who Controls AI Now — by Nathan House
🔓 Fable 5 Came Back. With More Locks.
After 19 days offline, Anthropic redeployed Claude Fable 5 on 1 July — the US lifted the export controls on 30 June. It's the same model, plus a new cybersecurity classifier that blocks the reported jailbreak over 99% of the time. The catch: it now flags more benign coding and debugging requests too, bouncing them to Opus 4.8. So the model that already refused half your normal work refuses a bit more of it. Community reaction is calling it nerfed; Anthropic says capability is unchanged when a request actually gets through. And the window's short — paid plans keep Fable until 7 July, then it's usage credits on top of your plan. Load your big jobs now.
- Anthropic — Redeploying Fable 5
- MarkTechPost — Anthropic redeploys Claude Fable 5
- On why these models refuse and stumble even as everyone adopts them: LLM Limitations: Why the Flaws Won't Stop Adoption — by Nathan House
🏛️ OpenAI Floats Handing the US Government 5% of Itself
Per a Financial Times scoop, OpenAI is in early talks to give the US government a 5% stake — about $42.6 billion against its ~$850B valuation. Sam Altman's pitch: give the public a financial stake in the upside, and ease the pressure from Washington. The obvious problem: the government is also the regulator. Hand the referee equity in the team and the incentive to wave models through gets a dollar figure attached. It's a floated proposal, not a done deal — no OpenAI newsroom post, just FT sourcing. But it's a strange direction for "safety" to point.
🎭 Claude Sonnet 5 Arrived — and Anthropic Barely Mentioned It
Normally a new frontier-lab model leads the week. This one got buried under Fable and GPT-5.6. Anthropic's own system card, page two, says the quiet part out loud: Sonnet 5 is "our most capable Sonnet-class model" but "not advancing the capability frontier." Below Opus on coding, reasoning and computer use; a bit better on knowledge work; fewer undesirable behaviours than 4.6. The point is price — $2 in / $10 out through 31 August, then $3/$15. One asterisk: a new tokenizer produces ~30% more tokens per request, so the effective cost creeps back up. Cheaper model, longer bill.
- Anthropic — Claude Sonnet 5
- Anthropic — Sonnet 5 system card
- When the gated frontier locks you out, the open-weight models are the fallback — and their own risk: Open Weight Models: A Security Guide to the Dangers — by Nathan House
🍌 Google Shipped a Fast Lane for Images and Video
Three drops. Nano Banana 2 Lite — the fastest, cheapest image model Google's made: ~4 seconds an image, about 3.4¢ per thousand, rolling into Search, the Gemini app, NotebookLM and Photos. Gemini Omni Flash — video generation and editing via the API at 10¢ per second of output. And NotebookLM can now spit out 60-second vertical video overviews. The theme is obvious: quality was last year's fight. This year it's speed and cost-per-thousand.
- Google — Build with Nano Banana 2
- Ars Technica — Nano Banana 2 Lite is Google's fastest, cheapest yet
🔬 Anthropic Built Claude Code — for Scientists
New app: Claude Science, an AI workbench for researchers. Not a new model — it runs Opus 4.8 as the engine and wraps it in scientific workflows. Think Claude Code, pointed at lab work instead of a codebase. Beta, macOS and Linux, on the paid plans. The interesting move isn't the model; it's Anthropic deciding the moat is the harness, not the weights.
📱 Cursor Put Your Coding Agent in Your Pocket
Cursor shipped a native iOS app. Launch always-on cloud agents, or remote-control the ones running on your desktop, from your phone. The workflow everyone's drifting toward: kick off the agent, walk away, check it from the sofa. Coding as something you supervise, not something you sit at.
🔗 X Now Hosts Its Own MCP Server
X stood up a hosted MCP server, so tools like Claude and Cursor connect to the X API with your own account permissions — no self-hosted plumbing. Read-only for now. Every platform's racing to become an agent-callable endpoint; X just made itself one with a click instead of a weekend of setup.
📝 Google Meet Will Take Your Notes Now
Official name: "Take notes for me." Gemini transcribes the call, writes a summary with action items, saves it to a Doc, and emails you after. AI Pro and Ultra subscribers. Granola and its clones built a business on exactly this — now it's native and free-with-plan inside Meet. The standalone-note-taker market just got its eviction notice.
✨ Gemini Spark Lands on the Mac — Plus a Codex Macro Pad
Two quick ones. Gemini Spark — Google's answer to OpenClaw — is now a macOS app for Ultra subscribers: it sorts your files, builds sheets from invoices, takes actions on your machine. And OpenAI teased Codex Micro, a programmable macro pad built with Work Louder (13 keys, joystick, rotary dial) for driving Codex. Full reveal 15 July. The agents moved from the cloud onto your desktop; now they're getting dedicated hardware buttons.
