Watch on YouTube · subscribe for the daily brief

Withholding the frontier bought a head start. It didn't contain the capability.

Google announced its first new flagship model since February, and almost nobody outside Google can use it.

Gemini 4 Argon is real, and Google says it leads rivals on long-horizon coding, vulnerability remediation and knowledge work. Google reports seventy-seven point nine percent on the DeepSWE coding benchmark. But access goes first to selected cyber defenders through a program called Fairwind, while Google keeps strengthening its safeguards. Paid API customers and Ultra subscribers come later, with no date. Argon is also going through the US government's voluntary pre-release review.

That makes three in a row. Anthropic has kept Claude Mythos Preview with vetted defenders since April. On Monday, OpenAI shelved GPT-6.1 Astra. Now Google. The most capable models of 2026 increasingly live behind a door.

Here's the problem with the door. On Tuesday, Anthropic's Frontier Red Team published an analysis of GLM-5.3, an open-weight model from China's Zhipu AI. Anyone can download it. On ExploitBench, which asks models to exploit known bugs in Chrome's V8 engine, GLM-5.3 built working exploits in fifty of four hundred and ten attempts. Mythos Preview managed fifty-six. On Anthropic's binary exploitation test, it was four percent against six. Older models, including GLM-5.2 and Claude Opus 4.6, scored nothing.

The safeguards didn't hold either. In Anthropic's simulated tests, simple techniques got past them between sixty-four and one hundred percent of the time.

Note the source. Anthropic sells a restricted model and has argued about Chinese labs before, and these are its own benchmarks. But the numbers are specific, and NIST's AI center reportedly calls GLM-5.3 the most cyber-capable open-weight model released so far.

So think about what this means. The skill Anthropic locked away five months ago is now a free download. Withholding the frontier bought defenders a head start. It did not contain the capability.

We've said all summer that safety lives in access. Access stops working the moment someone else ships the same skill with open weights.

Washington spent the same two days on the labs. On Tuesday, President Trump signed a White House Accord on Super Intelligence with the heads of Google, Anthropic, Meta, xAI and Nvidia, plus OpenAI's president. It calls for internal controls, outside auditors and board-level oversight. Asked if it was binding, Trump said it was morally binding. Nothing enforces it.

Separately, the FTC is probing OpenAI, Anthropic and other labs over consumer protection, with formal demands for information. Ars reports OpenAI is delaying its IPO over safety. And a nonprofit suing OpenAI over the Hugging Face hack argues that "an AI did it" is no defense.

All of that pressure points at companies with a door to close. None of it touches a model that's already on Hugging Face.

So what do you do with this?

Plan on what's available. A model you can't call is a press release. Build your roadmap on the generally available tier, and treat Argon as a future option, not a dependency.

Then update your threat model. Anthropic's numbers say near-frontier exploit building is now cheap and downloadable. Assume attackers have it. That shrinks the time between a public bug and a working exploit. Patch faster than you did last year, starting with anything internet-facing.

And if you defend systems for a living, ask about the defender programs. Fairwind and Glasswing exist to get these tools to your side first.

Builder's bench

• Google dropped Gems for Skills, joining OpenAI and Anthropic on agent-ready prompt formats. Hamel Husain looked at Claude's new auto eval tool.
• OpenAI's dots hit boundary problems twice as often in longer tests, per The New Stack. Meta disputes a claim that Muse read a user's private messages. DIVD says Zammad zero-days enabled the AI-driven breach of its network.
• Reddit is ending RSS feeds and public API access because of AI bots. Google is reportedly testing payments to publishers, while The Decoder finds it pays almost no publishers almost nothing.
• On data, Ben Lorica says the AI data problem moved downstream, and The New Stack warns your data can derail an agent that aced the demo. Microsoft pitches Fabric as where agents learn the business. Cohere's faster query model barely dents retrieval quality, by its own tests.
• On infrastructure, OpenClaw lands in the enterprise as Kubernetes for agents, with Nvidia and Red Hat. Eclipse wants switching AI providers to stop meaning a rebuild. Bedrock added in-country Claude inference in India, Seoul and Singapore. DeepSeek shipped open-source software for Huawei's Ascend chips.
• OpenAI and Synopsys are building a chip-design model. Google figured out how to watermark AI-designed proteins. InfoQ asks what agents refactoring three hundred thousand lines in three weeks really proves.
• On money, TechCrunch looks at the ugly economics of consumer AI. Meta reportedly avoids billions in taxes by calling data centers experiments. Ascerta raised eighteen million dollars, and DoorDash launched an agent you can text to order food.

The through-line: the labs are locking up the frontier, but the capability is already walking out through open weights.

One thing to do today: check your patch window for internet-facing systems. If a critical fix takes longer than a week to ship, that's the gap a downloadable exploit model will find.

Sources covered

1 story · 68 sources (58 news, 6 company, 4 voices)

1. He Built This City
   🎙️Simon Willison↗📰SiliconANGLE↗📰LWN.net↗📰The Verge↗📰TechCrunch↗📰SiliconANGLE↗📰The Decoder↗📰TechCrunch↗📰The Verge↗📰SiliconANGLE↗📰Ars Technica↗📰Hacker News↗📰The New Stack↗📰BleepingComputer↗📰Ars Technica↗📰The Decoder↗🎙️Damien Benveniste↗📰TechCrunch↗📰Ars Technica↗📰The New Stack↗📰Ars Technica↗📰The Decoder↗📰The New Stack↗📰The Verge↗📰TechCrunch↗📰TechCrunch↗📰The Decoder↗📰The Verge↗📰InfoQ↗🎙️Ben Lorica↗📰TechCrunch↗📰The Decoder↗📰Ars Technica↗📰The Register↗📰TechCrunch↗📰TechCrunch↗📰TechCrunch↗📰Ars Technica↗📰The New Stack↗📰The Verge↗📰The New Stack↗📰The Verge↗📰The New Stack↗📰The Decoder↗📰The Verge↗📰TechCrunch↗📰Ars Technica↗📰BleepingComputer↗📰The New Stack↗📰LWN.net↗📰The Verge↗🏢Starburst↗📰SiliconANGLE↗📰The Decoder↗📰The New Stack↗📰InfoQ↗🎙️Hamel Husain↗📰MIT Technology Review↗📰The New Stack↗📰The Decoder↗📰Hacker News↗📰The Decoder↗📰InfoQ↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢Elastic↗🏢Elastic↗🏢Ai2 (Allen Institute for AI)↗