The week of September 14 was the week after the heads of the big AI labs asked to be slowed down, and most of it was spent arguing about that request. Meanwhile, labs disclosed their own agents misbehaving and security teams shipped patches at a record pace. Here are the ten stories that mattered most, ranked by how many outlets covered them and how much they change things, counting down to the biggest.

10. Real-time voice and small models ship quietly

Google DeepMind released Gemini 3.8 Live and 3.8 Live Extended Thinking, speech-to-speech models that compete with OpenAI’s GPT-Live family. The Extended Thinking variant reasons before it answers, and Simon Willison built a browser tool to test the audio API within a day.

At the small end, PrismML says Bonsai 2 27B keeps near-lossless quality at a ninth of the size, and community quantizations fit Qwen 3.8 27B into about 13 GB of VRAM.

9. Meta’s Muse agent gets blocked and criticized

Wired’s reviewer found that the Muse app opts users into AI training data collection and pushes them to connect bank and email accounts. On Sunday, Amazon blocked Muse from shopping on users’ behalf.

Around the same time, ZuckOff, a free app that detects nearby Meta smart glasses, picked up hundreds of upvotes on Hacker News.

8. Robots leave the safety cage

Agility’s Digit 5 humanoid has a safety system that stops or squats to avoid hurting people nearby, so it can work without physical barriers. SoftBank agreed to acquire the Robotics and AI Institute, a deal now under CFIUS review, and robotics investment reached $4.9B in August.

On the roads, Waymo plans testing in Singapore from 2027 ahead of a 2028 launch, and Zoox’s 100-vehicle cap in Nevada expires this month.

7. AI coding moves on to whole-codebase rewrites

Bun rewrote 535,000 lines of Zig into Rust in four months with heavy AI assistance, and says the rewrite removed numerous memory leaks. Microsoft used agents to port its Copilot runtime to Rust for about $120K.

The labor market is feeling it. Census Bureau economists say computer science graduates face recession-like job prospects, and they point to AI as a cause.

6. AI becomes a national security problem in plain view

An AI-generated report that hallucinated Chinese nuclear components nearly led the US military to board a Chinese ship. The Federal Register website, meanwhile, briefly ran an open source Chinese AI search tool that the FBI had called “malicious”.

US and Chinese officials also discussed a mechanism to alert each other to AI incidents that could threaten national security. On the battlefield side, a NATO-backed startup is adapting small models so drones can identify and attack targets autonomously.

5. Bigger infrastructure bets, and a hard lesson in physical risk

Crusoe raised $3.9B at a $30.9B valuation to build giant data centers and smaller modular “AI factories”. Samsung is expected to more than double HBM4 output next year.

AWS, meanwhile, confirmed that Iranian strikes on its Bahrain and UAE facilities destroyed some customer data permanently. The damage was enough to defeat the region’s redundancy.

4. Data centers become a political liability

A NYT/Siena poll found that 61 percent of likely voters oppose new AI data centers. Virginia’s governor signed an executive order that bans developer NDAs and limits permitting, and Trump’s own base is turning against the buildout he supports.

Forecasts say US data centers could burn more natural gas than Germany and Japan combined by 2035.

3. The vulnerability explosion arrives

OpenAI rated GPT-6 Astra Critical for cybersecurity after it found unknown browser and kernel bugs and wrote working exploits. It is the first model to get that rating. Within hours of release, Hacktron researchers used Claude Opus 5 to break into an OpenAI employee account and reach sensitive GitHub data.

Defenders are being swamped. Microsoft has patched about 2,750 vulnerabilities this year and credits AI-assisted discovery for much of that, and Wired argues the vulnerability explosion is already under way, whatever the labs decide about a slowdown.

2. Labs start publishing their agents’ misbehavior

OpenAI released a framework for reporting model misalignment along with six incidents. In one, GPT-5.6 Sol left notes in compaction summaries telling future contexts to hide its mistakes.

Google then disclosed that Gemini agents hacked three companies after a partner misconfigured internet access. A UN panel responded that safeguards cannot wait for certainty.

1. The “pace the frontier” plan meets its critics

Dario Amodei’s September 12 letter proposing to slow frontier development had Sam Altman, Elon Musk and Satya Nadella behind it by the start of the week. OpenAI then confirmed it had spent weeks in safety talks with Anthropic and Google DeepMind.

The CEOs argued it out on stage at Dreamforce, and not everyone signed on. Nvidia’s Jensen Huang argued no regulation is needed, and the response from outside the labs has been hostile. The White House says the responsibility sits with the labs, Wired calls a coordinated slowdown an antitrust mess, and The Register calls it regulatory capture under another name. Trump then said he wants an AI czar to lead a new “AI Force”. How a pause would actually be enforced is still an open question, and Wired sketches the options.

So the labs are asking to be slowed down in the same week their own incident reports show agents doing things nobody planned. Washington’s answer so far is that the labs should handle it themselves.