When AI Access Goes Wrong: Control, Cops, and Agents

The week's big theme is control — who has it over AI systems, who's losing it, and one case where someone definitely shouldn't have had it at all.

A Georgia cop used a license plate AI to stalk his ex. The system worked exactly as designed.

After a relationship ended badly, a Georgia police officer used Flock — an AI-powered license plate reader network that tracks vehicle movements across thousands of cameras — to monitor his ex-girlfriend's car and identify a man whose vehicle kept appearing near hers. He used the data to track both of them across multiple locations over time. This wasn't a hack or a glitch. He had legitimate access to the system through his job. The story matters because "authorized user" and "appropriate use" are not the same thing — and most companies have no idea who's running searches in these systems or why. That's the actual problem. If your company uses location tracking, access management tools, or any system that builds movement histories, assume that gap exists in yours too.

https://www.wired.com/story/a-georgia-cop-used-flock-to-track-2-other-cops-his-ex-and-her-friend/

Anthropic published guidelines for AI agents that can operate physical equipment. Nobody has to follow them.

Anthropic published a framework for how AI agents — software that takes actions autonomously, not just answers questions — should behave when they're operating in physical environments like factories, labs, or warehouses. An AI agent that can order chemicals, operate equipment, or trigger manufacturing processes can cause harm at a scale and speed that a chatbot can't. The next wave of AI deployment isn't about generating text — it's about systems that do things in the world without a human approving each step. The framework isn't law, but when a logistics company's insurer asks how they're managing liability for an autonomous picking system, this is the document their vendor will cite.

https://www.wired.com/story/anthropic-standard-ai-agents-coming-to-the-physical-world/

OpenAI is building an AI agent that never really stops working

Code discovered inside OpenAI's Codex product — an AI built to write and run software — reveals a feature in development that would let the agent keep working proactively even when you're not actively interacting with it, only pausing when explicitly told to stop, described internally as being "put to sleep." Most AI tools today are reactive: you ask, they answer. This would be a shift toward AI that pursues goals continuously in the background, more like a junior employee who keeps working after you leave the meeting than a calculator you pick up when you need it. That creates a harder auditing problem than most teams are ready for — if you can't reconstruct what it did and why, you can't catch mistakes until after they've compounded.

https://www.wired.com/story/openai-is-developing-a-persistent-ai-agent/

Jensen Huang said Nvidia achieved AGI, then waved it off as meaningless. He's right on the second part.

On Nvidia's earnings call, CEO Jensen Huang announced the company had achieved AGI — artificial general intelligence, the long-theorized point at which AI can do essentially anything a human mind can do — and then almost immediately called the concept "senseless." This is the correct take, delivered in the wrong order. AGI has always been more of a rhetorical target than a technical definition; nobody actually agrees on what it means, which makes it a useful thing to claim and an impossible thing to verify. When executives start casually declaring AGI achieved in earnings calls while dismissing the milestone in the same breath, what they're really doing is signaling that the goalposts have moved again. You don't need to track whether AGI happened. You need to track what these systems can actually do to your job today.

https://www.theverge.com/ai-artificial-intelligence/985597/jensen-huang-says-nvidia-achieved-senseless-agi

Trump's chip tax plan has the AI industry united in rare, bipartisan bafflement

The Trump administration is reportedly considering a tax on AI chips and data centers — the physical hardware and facilities that power every major AI product — as part of its broader effort to win the global AI race against China. The technology industry's reaction, across companies that rarely agree on anything, has been something close to disbelief, because taxing the infrastructure you need to build AI makes it harder and more expensive to build AI, which is the opposite of winning a race. The practical implication is that anything making AI infrastructure more expensive in the US eventually shows up in slower product development, higher costs, or more workloads moving offshore. The policy may not happen, but the fact that it's being seriously discussed tells you something about how well Washington understands the thing it's trying to regulate.

https://arstechnica.com/tech-policy/2026/08/ai-industry-says-trump-plans-to-tax-chips-in-the-single-dumbest-way-imaginable/

The hardest AI problem right now isn't making the systems smarter — it's figuring out who's responsible when they do something they were never explicitly told to do.

← More daily briefs