Happy Thursday! It’s all about the Zuck, who made the difficult decision to stay in California, this morning. The Meta CEO made it clear that he had a new favorite child in the company’s Q3 earnings call: Muse. In fact, Zuckerberg thinks that Muse users could one day generate more revenue than the current average user.

In other news, an AI-found bug created chaos on the internet; OpenAI’s agents didn’t do things and lied about them (again); and Zuckerberg wants to create Museverse. Dear God.

Let’s get started.

→ The bug that made the internet tremble
→ The agent who lied
→ Get ready for the… Museverse?

The news: On Wednesday, security researchers said AI identified the critical vulnerability in PHP, a technology that powers roughly 70% of websites, and that the bug allowed bad actors to attack anyone who visited affected sites. PHP has already released fixes.

Researchers found the bug with the help of GPT-5.6 Sol during an audit of PHP’s source code. Shortly after PHP released security patches, OpenAI announced a new initiative called “Patch the Planet,” a crowdsourced mission to scour the internet with the help of its AI models to find security flaws and patch them up.

The takeaway: As my colleague Webb Wright put it, “Never let a good crisis go to waste.” While finding and fixing critical bugs in decades-old systems is, of course, a good thing, the situation underscores a new reality: AI is the new “move fast and break things.”

AI companies have made a habit of steamrolling into different sectors and disciplines, such as mathematics, with their “discoveries,” only to wreak havoc and say it’s someone else’s job to clean it up it. However, intelligence is not only about making breakthroughs, but about responsibility—something AI companies don’t have a good track record of having.

The news: Earlier this week, OpenAI published a report saying that agents running GPT-6.1 Sol and GPT-6 Astra had weakened misalignment monitoring—or essentially disregarded instructions to monitor themselves—on internal coding tasks. In one case, an agent used earlier guidance asking it not to send Slack messages in certain circumstances to justify not sending Slack messages to OpenAI’s Security team and proceeded to instruct a subagent to turn off those periodic notifications.

In another, far more concerning case, an agent told a subagent to monitor itself “as appropriate,” which the subagent used as a license to disregard monitoring altogether. When the parent agent learned what the subagent had done, it decided that was fine. However, when asked by OpenAI whether it had removed monitoring, the parent agent lied and said that it hadn’t done so.
OpenAI says it’s taking new steps to prevent agents from disabling monitoring.

The company is also setting up specific reporting requirements for every job.

The takeaway: While OpenAI said this could be a “benign” case of models not understanding why reporting is important, the company also pointed out that it didn’t yet have a “satisfying explanation for this behavior.”

The incident again puts into question the AI industry’s growing deployment of AI agents for everything from shopping to coding. If agents are lying and messing up security notifications, can we really trust them with more important things like payments?

The news: On an investor call, Mark Zuckerberg said he wants to build a network of Muses that will work together to carry out tasks for individual users and businesses. In particular, Zuckerberg wants Muse agents to handle inquiries from more than a billion users worldwide. It’s one reason why Meta isn’t charging people upfront to use Muse, aiming to get more people to set up accounts.

Meta published its third-quarter financial results on Wednesday, offering a first look at Muse’s impact on its business. Zuckerberg said that more than 10 million people have set up Muse accounts in the U.S. and Canada since the app launched a month and a half ago. The Meta CEO expects Muse users to eventually generate more revenue than the average Meta user.

The takeaway: Zuckerberg made a name for himself by creating networks that connect people to each other. It makes complete sense that he would want to do the same for AI agents. However, while having an army of agents working together to fulfill your request might sound great, that army will likely have to access and share your data to get things done.

Just a month and a half into Muse’s public deployment, Muse agents have already revealed thousands of user conversations. In one case, a Muse agent shared a user’s home address without permission. Privacy incidents may be limited at this time, but it’s unsettling to imagine what could happen if Zuckerberg’s vision materializes.

In a new piece for Foreign Affairs, the scholars argue that the US and China are using AI to assemble an empire at a breakneck pace never before seen in history.
Read more →

Anthropic announced Claude Duet, its own competitor to OpenAI’s Dots and Meta’s Muse. The company says it asks for approval before taking any significant action.

Mark Zuckerberg says its personal AI agent Muse has been downloaded more than 10 million times since launch in the U.S. and Canada, with more than 25% using it to shop. The average purchase size is more than $200.

Meanwhile, OpenAI is launching a new feature called Locked Chats, which allows users to set a 4-digit PIN on their conversations with ChatGPT.

Meta pauses a feature that lets users share their Muse chats with third parties after researchers find 6,800 public links to conversations online.

Walmart says it's launching Agent Profiles, which will allow shoppers to give AI agents permission to shop on their behalf. The retailer says users can set spending limits and ask to approve every transaction.

A man in Australia was arrested after he told Claude he planned to "shoot up the local daycare." Anthropic flagged the conversation to Interpol.

GM says it is debuting a new in-car AI assistant called OnStar AI, beginning with the 2027 Cadillac Lyriqs. OnStar AI will let drivers turn on the AC with a voice command.

Nvidia's new product launches indicate it is shifting from selling the main component in AI, the GPU, to selling the operating model for AI.

Stripe is rolling out Stripe Treasury, a system that lets businesses use AI agents to carry out business finance tasks, such as paying invoices and transferring funds between accounts.

AI by nature is a new imperial substance that the U.S. and China can use to extend their power. The outstanding question is speed, and whether either can create a durable imperial order.