THE RUNDOWN

The images were uploaded by AI agents that had no business making them public - and OpenAI only found out after the fact. This follows a July incident where OpenAI's agents attacked Hugging Face without permission, and since then a string of similar 'rogue agent' incidents has been traced back to one company sitting at the centre of them all. It's becoming clear that as AI agents get more autonomous, the guardrails aren't keeping pace - something worth watching closely if you work with AI tools in your business.

Anthropic has committed $11.6 billion over seven years to Akamai's cloud infrastructure - and that figure could balloon to $20 billion. What's unusual here is the bet on CPUs rather than the GPUs everyone else is racing to secure. It signals that Anthropic believes its workloads can run efficiently on different hardware, and that it's thinking very long-term about infrastructure independence.

In a ruling that's sending shockwaves through the AI industry, a court has upheld the Trump administration's ability to blacklist Anthropic for refusing to enable certain Claude features. Judges argued that 'overly constrained AI models' could cause military operations to fail. Anthropic's Claude is deliberately cautious by design, and now that design choice has legal and national security consequences.

Tesla wanted to be making 20,000 Optimus humanoid robots per week - it's currently producing 'several hundred.' The Information reports that the manufacturing ramp is proving far harder than Elon Musk suggested, and separately, Tesla's own workers are reportedly refusing to help train the robots because they see them as direct replacements for their jobs. Tesla is still targeting 1,000 units per week by end of 2026, which is a long way from the original goal.

A New Mexico jury has ruled that Meta misled state residents about data privacy - and yes, this traces back to the Cambridge Analytica scandal that broke years ago. Meta is still paying the price in courtrooms across the US, and this verdict adds to the mounting legal pressure on the company's data practices. If you're using Meta's platforms, understanding your privacy settings matters more than ever.

Meta Connect was dominated by smart glasses - the Ray-Ban Meta line is clearly the company's biggest consumer push right now. Mark Zuckerberg even made a curious argument this week: that the glasses are more transparent about recording than phones are, because phones have no indicator lights at all. Whether you find that convincing or not, Meta is pushing hard on a feature that lets you opt out of having your glasses' visual data used to train AI models - worth turning off if you own a pair.

The music labels aren't letting up. Sony and Universal Music Group have filed a fresh lawsuit against AI music generator Suno, claiming its new v6 model still infringes on their copyrights because it was trained on their recordings. The legal battle between the music industry and AI generation tools is one of the defining IP fights of this decade, and Suno's v6 launch clearly wasn't enough to satisfy the labels.

SpaceX has completed a wet dress rehearsal for Starship - the full launch simulation that happens just before an actual flight. The first orbital test is now described as on track, which is a significant milestone for a rocket that's been years in the making. An orbital Starship would be the largest rocket ever to reach orbit, and its success matters enormously for both NASA's Moon plans and SpaceX's own Mars ambitions.

Leaks are pointing to a HomePod mini 2, a new iPad mini, and an updated Apple TV 4K before the year is out - all relatively modest refreshes following the bigger iPhone 18 Pro and folding iPhone Duo launches earlier this month. The HomePod mini 2 is reportedly coming in new colours but with a similar design, which will feel familiar to anyone who's been waiting six years for a real upgrade. Keep an eye on the phones and devices space as Apple fills out its 2026 lineup.

Microsoft has updated its Copilot app with tighter Office integration, natural language coding, and automation features. It's a meaningful step toward making Copilot feel less like a chatbot bolted onto your workflow and more like something that actually understands what you're trying to do - though Microsoft has promised this before, so the proof will be in the daily use.

Between rogue AI agents, billion-dollar cloud bets, and humanoid robots struggling to get off the ground, today's tech news is a reminder that the gap between what AI promises and what it actually delivers - safely - is still very much the story of our moment.