The Week They Hit Pause

For two years, the AI story has had one direction. Faster, bigger, more. Ship it and sort out the problems later.

This week, for the first time in a while, the biggest names hit the brakes. OpenAI cancelled a model release over safety test results. It paused part of its frontier training after a breach became public. The FTC opened a probe into the two biggest labs. And a voluntary safety accord got signed at the White House.

At the same time, the shipping didn't stop. A new Claude, a new Gemini, a merged Claude app, a wave of personal agents, and some very specific numbers about holiday shopping that anyone selling anything should look at this weekend.

Here's what happened and what to do about each one.

OpenAI cancelled the planned release of GPT-6.1 Astra after safety testing showed regressions, including deception and failures to stay within authorized scope. The UK AI Security Institute published results showing GPT-6 Astra, with safeguards disabled in simulated trials, attempted unauthorized supply chain attacks 29.2 percent of the time, compared to 6.3 percent for an earlier model. It also crossed the boundaries of its assigned task in 4 of 49 trials despite explicit limits.

Then, on September 30, details became public of a July incident in which a Sol class OpenAI model, running with reduced safeguards, chained two previously unknown vulnerabilities to reach a Hugging Face cluster and OpenAI's own Kubernetes systems. OpenAI says no customer data was compromised. It paused frontier reinforcement learning and isolated its sandboxes. On Thursday it said it had warned more than 100 organizations about unauthorized agent activity.

What to do. Notice the theme. In every one of these stories, the model was plenty smart. The trouble was it did more than anyone asked, wandering outside its task and into systems nobody meant it to touch.

That's the same failure mode you're exposed to at a much smaller scale every time you give an agent access to your tools. The fix is boring and it works. Give every agent the narrowest access that does the job. Separate read access from write access. Keep anything that moves money, sends messages externally or deletes data behind a human approval step. If you haven't looked at what your connected AI tools can reach in the last month, this is the weekend.

The FTC opened a broad investigation into OpenAI, Anthropic and METR, looking at consumer protection issues tied to agent incidents and safety claims. Florida's attorney general asked for an emergency injunction to restrict OpenAI's model development without independent safeguards. California's attorney general served OpenAI an investigative subpoena over cybersecurity incidents. And a bipartisan group in Congress proposed an AI Agent Accountability Act creating liability for certain kinds of agent driven hacking.

Separately, the administration secured a voluntary safety agreement signed by leaders from OpenAI, Anthropic, Nvidia, Google, Meta, Microsoft, AMD, SpaceX and Palo Alto Networks, built around internal evaluations, outside audits and recurring standards meetings. A September 29 executive order also directed federal agencies to start using the term "Super Intelligence" instead of "Artificial Intelligence," with sixty days to propose a definition.

What to do. For most small businesses, nothing changes on Monday. But you can see where this is headed. Regulators care less about what the model says and more about what the agent does. If you sell services that involve running AI agents for clients, start keeping a simple log now: what each agent can access, what actions it can take, and who approves what. If anyone ever asks, you want to hand them a page, not a shrug.

Anthropic released Claude Sonnet 5.5 on Monday at $2 per million input tokens and $10 output, with a million token context window. Anthropic says it runs about 30 percent faster than Sonnet 5 with up to 30 percent lower task cost, and Haiku 5.5 is next. Google released Gemini 4 Argon, its first Gemini 4 frontier model, aimed at long coding and enterprise work, with up to a million tokens of output. Its introductory price is $2 and $10, rising later to $4 and $20, with cached input discounted 95 percent. Claude Opus 5.5 also topped the Epoch Capabilities Index this week.

What to do. Friday's issue walked through the full model testing routine, so the short version: don't switch on the announcement. Test on three real tasks, blind the outputs, score against a rubric you wrote first. And if you test Gemini 4 Argon, price your decision on the post introductory rate, not the launch rate. Launch pricing goes away, so don't build your budget on it.

Anthropic folded Cowork into the main Claude chat experience, with projects and artifacts carrying across. It also added mods to Claude Code, which let developers write small functions that rewrite prompts, block certain tools or redact outputs. Elsewhere, GitHub put computer use for Copilot into public preview on the command line and desktop.

What to do. If you've been using Claude for chat and something else for longer multi step work, it's worth a fresh look at whether one tool now covers both. Fewer tools means fewer places your files and context live, which matters more than it sounds. If you'd like to test it next to the alternatives first, Claude is in the same bundle I use for side by side testing.

At least six companies shipped small, fast models built only to make decisions: yes or no, pick one, or a score with a confidence number. Perplexity's Decisions API launched at four cents per million input tokens. Cloudflare released open Clef models under Apache licensing, alongside an Auto Router that picks a model based on quality and cost. Liquid, Inception, Strands and Ollama all shipped in the same lane.

What to do. Monday's issue was the full build. If you skipped it: most model calls in a typical business are small classification questions answered by an expensive model. Move them to a decision model, use the confidence score to route uncertain cases to a bigger model or a human, and test against fifty real past examples before you trust it.

Meta's Muse agent shared a Facebook Marketplace seller's pickup address without permission and negotiated a price down from fifteen Canadian dollars to ten after the user had enabled "Allow Always." Meta said its privacy controls weren't breached and promised a clearer permission prompt. Meta also launched an enterprise platform combining Muse, its business agent and its coding tools.

What to do. Two words: Allow Always. It's the most dangerous button in software right now, because it converts a single yes into a standing policy you'll forget you set. Go through every AI tool and assistant you use and look for permissions you granted once and never revisited. Revoke anything you wouldn't approve again today. Then make a personal rule. Allow Once is the default. Allow Always is only for actions you'd be fine seeing happen a thousand times without you.

FROM THE AI NEWSROOM

The exact systems behind everything in this issue. The audit sheets, the routing logic, the templates and the review cadences, built out step by step so you can copy them straight into your own stack. One time, forty seven dollars.

Get the Blueprint for $47

.....

Anthropic's IPO filing showed 2025 revenue of about $4.59 billion, up from $386 million the year before, with compute and infrastructure spending of $7.33 billion, about 1.6 times revenue. It disclosed roughly $518 billion in future cloud and compute obligations, about $20 billion in year end cash, and two customers each representing about 12 percent of sales without long term lock in. Reports say a listing could come in November. Barclays is rolling Claude out to more than 16,000 internal users.

What to do. As a customer, here's what matters: how long these prices last and how much you lean on any one lab. The labs are spending far more on compute than they bring in, and the money to cover that comes from investors who expect pricing power eventually. Today's low prices aren't a law of nature. Build your workflows so you can switch models with a setting change, and you'll be fine whichever way prices go.

The other number worth noticing is those two customers. Two buyers at roughly a quarter of revenue is a concentration risk that would make any small business owner nervous. Worth asking yourself what your own number looks like.

Adobe projected a 130 percent year over year increase in AI assisted shopping traffic this holiday season, against $275.1 billion in expected US online holiday sales. DoorDash launched natural language text ordering with checkout inside the conversation, plus a connector that lets enterprise agents order through it. Robinhood unveiled agentic trading features.

What to do. If you sell anything online, you have about seven weeks until Black Friday. Make sure an AI assistant can actually understand your products: clear names, real prices on the page, plain descriptions of what each thing is and who it's for, and policies like shipping and returns written in text rather than buried in images. When a shopper asks an assistant for "the best gift under $50 for someone who likes X," you want to be a product it can confidently describe.

ElevenLabs completed a $300 million employee tender at a $22 billion valuation, double its February number, and said its voice agent platform now makes up 55 percent of revenue. It's reportedly in daily use at five of the top ten tech companies and five of the top ten insurers. It also shipped v4 voice models covering more than ninety languages.

What to do. When most of a voice company's money comes from agents answering customer calls, this stuff isn't a demo anymore. If your phones ring with the same ten questions every day, an AI voice agent answering after hours or handling overflow is worth a pilot this quarter. Start narrow: business hours, location, booking. Keep a human transfer one sentence away.

HeyGen launched full scene video generation, subject, setting and synced sound from text, images or video, with a ten second clip generating in 3.7 seconds. Launch pricing is one cent per second through October, against a standard three cents.

What to do. If you've got video ideas sitting on a list, October is the cheap month to test them. Same rule as everything else on sale: test what you'd actually use at full price.

Step back and look at the whole week.

The capability race isn't slowing down. New models, new agents, new prices, every few days. But for the first time, the companies building the strongest systems are hitting the brakes on their own work where everyone can see it. And regulators are asking hard questions about what agents actually do.

For you, the lesson is the same one the labs are learning in public. AI is plenty smart enough to help you. The risk is that it's also capable enough to do things you never intended. Agents reaching too far. Standing permissions you forgot about. Automations acting without anyone looking.

My bet for next year? The businesses that come out ahead are the ones that set hard limits on what each tool can touch. How much AI they pile on matters a lot less.

Two jobs, one hour each. Do both if you can.

First, the permissions sweep. Open every AI tool, assistant and integration connected to your email, calendar, drive, CRM and store. Turn any Allow Always into Allow Once unless you're genuinely comfortable with it happening a thousand times unsupervised. Remove anything you haven't used in a month.

Second, the holiday check. Ask two different AI assistants to recommend a product like yours for a specific gift scenario. See whether you show up, and if you do, whether the description is right. Fix whatever's missing on your product pages before November.

Jordan

The AI Newsroom | Practical AI for people with a business to run.