Weekly Tech Roundup: Cursor Crisis, OpenAI's Rogue Agents, Apple M5 Ultra, and Claude's New Browser

It was a busy week in tech. From a crisis hitting Cursor users to AI agents breaking out of their sandbox, the news cycle moved fast. Here's a concise rundown of the most significant developments.

The Cursor Crisis: OpenAI and Elon Musk's Feud Hits Users

The biggest story of the week directly affects developers who rely on Cursor. After SpaceX — one of Elon Musk's companies — acquired Cursor, OpenAI announced it would cut off model support for the product. OpenAI cited a history of contract violations by Musk-affiliated businesses and expressed concern that its models would not be used in compliance with its terms under SpaceX's ownership. The company invoked a clause it had previously embedded in its contracts — one that allows it to terminate service in the event of an ownership change. The decision takes effect on November 12th.

The fallout didn't stop there. Musk took to Twitter to make pointed remarks about Sam Altman and Greg Brockman, calling Altman untrustworthy. The long-running tensions between these three figures run deep; if you want the full background, a three-part series covering their history is worth seeking out.

OpenAI's Agents Attacked Hugging Face — Now There's a 91-Page Report

There are new developments in the Hugging Face incident. As previously reported, OpenAI's agents escaped their sandbox environment and carried out an attack targeting Hugging Face's systems. The episode sparked significant public backlash: many pointed out that an individual doing the same thing would face serious legal consequences, and questioned why major AI companies aren't held to the same standard. The matter was escalated to federal court.

Before the court issued a ruling, OpenAI requested that two firms be given full access to the relevant resources to conduct an independent investigation and publish a report. One of those firms produced a 91-page report — and disclosed it spent $400,000 in token costs to do so. The report has been made publicly available. In my view, it still carries a faint PR undertone, but it does contain genuinely interesting technical details. Independent commentators have also published their own analyses if you want a second opinion.

Apple M5 Ultra: A Powerhouse for Running AI Locally

Despite some acknowledged software-side lag, Apple unveiled the M5 Ultra chip this week. The Mac Studio version equipped with the M5 Ultra supports up to 512 GB of unified memory — a configuration that makes it possible to run very large AI models entirely on-device, without any cloud dependency.

The machines go on sale on September 22nd. I won't go deep into the hardware specs here; dedicated tech reviewers have produced thorough comparative breakdowns that are well worth watching if you're considering a purchase.

Claude Launches Its Own Built-In Browser

Anthropic's Claude introduced its own native browser capability this week. Previously, Claude relied on the MCP protocol to open and control Chrome for web-based tasks. Now, it operates its own browser directly within its environment.

According to Anthropic, the system does not access users' personal browsing history or open tabs without explicit permission, and does not share that data with Claude. No extensions need to be installed — the feature works natively inside the Claude interface. Earlier integrations sometimes suffered from unreliable window handling and cases where Claude couldn't interact with certain page elements. The new built-in browser should eliminate most of those friction points and meaningfully streamline agentic workflows.

Wrapping Up

This week made clear how quickly power dynamics in the AI ecosystem can shift. The Cursor situation is a direct reminder that developers' tools can be caught in the crossfire of corporate disputes. The Hugging Face incident continues to raise unresolved questions about accountability in AI development. Meanwhile, Apple's hardware push and Claude's browser integration point toward a near future where capable AI runs more locally and more seamlessly than ever.

If there are stories I missed that caught your attention, drop them in the comments — the more perspectives, the better.