Two weeks ago we told you about a Google model you couldn't use yet. Today you can. On the same day, an OpenAI engineer stood in front of a federal judge, accused of hiding a listening device inside one of the company's conference rooms.

On the surface the two stories don't connect. Look closer and they're about the same thing: AI labs have spent years guarding their model weights, and this week's headline threat came through a USB port behind a TV.

Here's the story, pieced together from the federal complaint and OpenAI's own account.

On September 30, during a routine security sweep, OpenAI found a device wrapped in toilet paper and plugged into the USB port behind a conference room TV. OpenAI calls it "sophisticated": built for stealth, with encrypted audio and encrypted data transfer.

What came next reads like a heist movie run backwards:

He was arrested Friday, October 9, charged with using or attempting to use a device to intercept oral communications, and fired. His first court appearance was today.

OpenAI says it has "no indications" he accessed or exfiltrated model weights, source code or customer data, and that "the FBI's investigation is ongoing." Nobody has said publicly why he did it or who, if anyone, it was for. OpenAI also says he was planning to plant more devices.

Nearly everything we know so far comes from OpenAI and the prosecution, and none of it has been tested in court. These are allegations, and he's presumed innocent.

Since late September, Argon has only been available to a hand-picked group of cyber defenders. As of today it's live for paying users and developers.

Who gets what:

What Google says is new: more direct answers with less hedging, progress updates during long tasks, a Tasks toggle that keeps working after you close your laptop, Antigravity's agentic coding built into the Gemini macOS app, and connections to 40+ apps, including Instacart, OpenTable, Adobe and Shopify.

API pricing (the part worth putting in your calendar):

On Google's own benchmarks, Argon scores 77.9% on DeepSWE against Claude Opus 5.5's 74.2%. On the independent Artificial Analysis index it sits at 53, behind Opus 5.5 at 58 and Muse Wonder at 56. So it wins Google's chart and trails on the neutral one, which is the pattern we covered in last week's five questions to ask any launch chart.

Google's war stories are hard to forget, though. It says Argon found critical bugs in Project Wing's drone simulator that had been sitting there since 2018.

For OpenAI: a lab that's worth hundreds of billions found its most expensive secrets potentially exposed through the cheapest attack there is, which is someone walking into a room. Weights sit behind layers of encryption, but a conversation in a meeting room is just air. Expect every frontier lab to quietly re-sweep its offices this week.

For you: Argon's intro pricing makes it one of the cheapest frontier-class models to test right now, and the price doubles on December 1. If you have evals to run, run them in the next seven weeks.

For everyone: in both stories, AI ended up as the security guard. Google built Argon for defenders first. OpenAI says a model helped catch an insider. The tools are turning inward.

The stories moving fast beyond today's headlines:

📝 Anthropic and OpenAI reportedly sought pre-publication control over evaluator reports: Reuters says both labs pushed for the right to review third-party safety findings before they go public.

✉️ OpenAI's three fired safety researchers publish a letter disputing the company's account: Korbak, Balesni and Wang push back point by point.

🌿 Anthropic reportedly built "Scion" to replace Git internally: the reported reason is agents producing around 30,000 merges a day, a volume Git was never designed for.

🏠 Apple's $299 HomeView won't launch in the EU: the AI Act is the reason given, and it's one more product that skips Europe at launch.

🛡️ Argon's 2018 bug find: if Google's account holds up, it's a solid public example of a model surfacing long-buried bugs in production code (details in the Superpower Daily link above).

🧪 Google AI Studio: the fastest way to put Argon through your own prompts before the price doubles (Ultra tier for now).

🚀 Google Antigravity: Google's agentic coding environment, now built into the Gemini macOS app.

💰 Gemini API pricing: bookmark it, because the December 1 jump will change your cost math.

Google spent weeks deciding who was trustworthy enough to use Argon, and OpenAI just learned the hard way that the person you need to worry about may already have a badge. Both labs are coming to the same conclusion: the danger isn't only what the model can do, it's who's standing near it. Weights get encrypted, but conversations, calendars and the USB port behind the TV usually don't. The lesson carries well beyond AI labs. Your threat model is probably full of firewalls and API keys, and almost none of it covers the room where people actually talk. This week the cheapest attack beat the most expensive defenses, and what caught it was a camera, a forensic image and a model reading through the evidence. Lock down the obvious stuff first, because that's where people get in.

Stay building. 🤖