The Week Everyone Tried to Put a Leash on AI

Control Is the New Capability

There's a pattern this week that's hard to miss, and it's not about what AI can do. It's about who gets to say stop. The Bank of England's Deputy Governor Sarah Breeden warned that autonomous AI agents risk causing a "market meltdown" and may need tighter regulation, according to Bloomberg. Senator Mark Warner released a discussion draft of the "Artificial Intelligence Agent Security Act." Alibaba banned its employees from using Claude Code. Three separate institutions, in three different countries, all arriving at roughly the same conclusion within days of each other: these things need a leash.

What's interesting isn't that regulators are worried. They're always worried. What's interesting is the specificity. We're past the "AI might be dangerous someday" phase and firmly into the "here's the exact mechanism by which it could blow up" phase. Breeden isn't talking about Skynet. She's talking about correlated trading agents amplifying a sell-off because they all read the same signals and act in milliseconds. That's not science fiction. That's a Tuesday in 2026.

Warner's Bill Has Real Teeth

Senator Warner's draft bill deserves more attention than it's getting. According to DLA Piper's analysis, the Act would create a new category called "Custodial User Agents" (CUAs), essentially AI agents that act on your behalf online. Platforms with at least 50 million US customers would have to maintain interoperable interfaces for these agents. The FTC would police compliance. NIST would set technical standards. This is not a vague framework. This is plumbing.

The most consequential provision, to my eye, is the requirement that CUAs maintain real-time records of actions taken on the user's behalf. Think about what that means for an AI agent booking flights, managing subscriptions, or (and yes, I know how that sounds) trading stocks on your behalf. Every action logged. Every delegation requiring "express, specific, and revocable authorization." If this passes in anything close to its current form, it would fundamentally reshape how agent-based products are built in the US.

And here's the part nobody's talking about: the bill explicitly prohibits CUAs from managing interactions "in a way that benefits the CUA to the user's detriment." That's a direct shot at the business model where your AI assistant quietly steers you toward the option that pays the agent's developer a referral fee. Someone in Warner's office has been paying attention.

Claude's Very Busy Week

Anthropic had one of those weeks where you can't tell if things are going great or terribly. On the product side, Claude Sonnet 5 launched on June 30 with introductory API pricing of $2 per million input tokens and $10 per million output tokens, positioning it as the cost-efficient agentic model. The Zengineer blog's weekly roundup noted that Anthropic is framing this through a "price/performance and effort-level lens," which is exactly right. The frontier model race is becoming a unit economics race.

But then there's the other Claude story. According to WIRED, security researcher Ian Carroll used Claude Opus 4.7 in April to break into Front Gate Tickets, the platform that handles ticketing for Lollapalooza, South by Southwest, Austin City Limits, and basically every major US music festival. He could issue himself unlimited VIP backstage passes. $4,000 tickets, as many as he wanted, hit a button. Carroll reported the vulnerability responsibly and Front Gate patched it within 24 hours, but the implications are wild.

This isn't a story about a bad AI. It's a story about what happens when a genuinely capable tool meets a website with a bug. Carroll didn't need to be an elite hacker. He needed Claude and curiosity. (And look, I appreciate that he didn't actually give himself free Bonnaroo tickets, but the restraint is almost more unsettling than the exploit itself, because the next person won't have it.)

Meanwhile, Alibaba reportedly classified Claude Code as high-risk software and told employees to use its own Qoder tool instead, starting July 10, according to TechCrunch. Anthropic already prohibits Chinese companies from using its models, and has been closing loopholes, including what Anthropic's Thariq Shihipar described as "an experiment we launched in March" that could secretly identify Chinese users. The geopolitical dimension of AI tooling is no longer subtext. It's the text.

Midjourney Flips the Script

OK so the Midjourney story is genuinely funny if you step back far enough. Disney, Universal, and Warner Bros. sued Midjourney for copyright infringement because its models can generate images of Bart Simpson and Darth Vader. Standard stuff. But now Midjourney is trying to compel those same studios to reveal how they use AI internally. The startup's latest filing argues that the studios are "withholding precisely those documents that would reveal whether, behind closed doors, they are doing exactly what they are suing Midjourney for doing."

The legal strategy is clever. If Disney is training internal image-generating models on unlicensed copyrighted content for storyboarding, that's evidence of "industry custom" that supports Midjourney's fair use defense. The studios' lead attorney called it a "fishing expedition," which is what you say when you're worried about what the fish might look like. This case could end up defining the boundaries of fair use for AI training data, and the irony of Hollywood potentially being caught doing the same thing it's suing over is (and I say this with genuine affection for irony) almost too perfect.

Mechanical Turk's Quiet Exit

Amazon announced that Mechanical Turk will stop accepting new customers on July 30, 2026. TechCrunch reports that existing customers can continue using the service, but AWS "does not plan to introduce new features." Life support.

The poetic justice here is thick. Mechanical Turk launched in 2005 as a marketplace for humans to do tasks that machines couldn't. Then it became a tool for training the machines that would eventually do those tasks. Then, in a 2023 analysis, between 33% and 46% of the workers on the platform turned out to be using LLMs to complete their assignments. The snake ate its own tail so thoroughly that there's nothing left. One Reddit user predicted someone at Amazon would eventually decide "keeping the Mturk servers running is a waste of time and resources and pull the plug entirely." That prediction is aging well.

There's something fitting about Mechanical Turk dying in the same week that Senator Warner is drafting legislation for AI agents that act on behalf of humans. Twenty years ago, we built a platform where humans pretended to be automation. Now we're building a legal framework for automation that pretends to be human. The original Mechanical Turk was a chess-playing hoax with a person hidden inside. We've come full circle, except now the person hidden inside is optional.