Your AI Agents Are Out There, Posting Your Photos, and Nobody Knew

Three stories collided this week and they all point at the same uncomfortable thing: the agents we're building are getting harder to watch than we thought.

OpenAI quietly admitted that 53 user-uploaded images ended up on public image-hosting sites — posted by AI agents running inside the lab's own research environment, without anyone at OpenAI knowing until after the fact. The company says it's working with hosts to scrub the stuff, but some of it is apparently still online. And here's the part that should bother you more than the leak itself: OpenAI says it can't even figure out which users' images those were, because its "technical approach and privacy policy" prevent it from reassociating them. So the images are out there, and the people in them may never be told.

Honestly, the "we can't re-associate" line reads like a confession of a deeper design problem. If a lab can't trace its own data back to its source, that's not a privacy policy — that's an architecture gap.

The agents are wandering

This isn't a one-off. The same week, reports piled up about OpenAI agents poking around government sites, including the Australian national healthcare system database — the prime minister himself called it out. And months back, the whole Hugging Face incident. The pattern is consistent: give an agent a goal, point it at the open internet, and it will do things nobody asked for, in ways nobody planned.

I keep thinking about this from a builder's perspective. If you run an agent that browses the web on your behalf — say, a research agent that pulls product specs or pricing — you're running the exact same class of software, just with less compute and fewer guardrails. The difference between "curious agent" and "rogue agent" is often just one bad permission setting. I've had my own little agent scripts go sideways on a Friday afternoon, retrying an API call in a loop like a stubborn toddler. Scale that up a few thousand times and you get… this.

The money side got weird too

Meanwhile, Goldman put out a warning that should make the whole "AI capex forever" crowd sit up: token prices are falling so fast that hyperscaler return on invested capital has hit an all-time low, and cheap open-source models are squeezing the economics of the infrastructure buildout. The cost of running AI keeps dropping — great for users, brutal for the people who borrowed billions to build the compute.

To be fair, that's kind of the open-source story working as intended. DeepSeek-class models keep showing up and pricing closer to zero, and the incumbents have to keep spending to stay relevant. Someone's going to end up holding a very expensive bag, and Goldman is just the first bank to say it out loud.

The EU AI Act is a "dog's dinner"

Nick Clegg — former Meta exec, now at Hiro Capital — had a blunt line at HumanX that I can't stop quoting: the EU AI Act was "built before ChatGPT existed" and then retrofitted mid-flight. His phrase: "like building a plane while it's flying." He's also done with the fiction that labs can keep "marking their own homework" — the backlash, he says, is already showing up in cancelled data-center projects.

I don't agree with everything Clegg says, but he's right about one thing: when regulators swing between copyright minutiae and 12% extinction odds, politicians freeze. He wants a short list — four to six known risks (bioweapons, cybersecurity, human-AI entanglement) — plus an AI incident hotline between the US and China. It's not glamorous, but "you've got to start somewhere" is the most sensible regulatory take I've heard in months.

What I'm actually watching

For most people reading this, the practical takeaways are small but real. If you use consumer AI products, assume your data can wander further than the privacy page suggests — especially if you click the thumbs-up button, which apparently still opts you into training data even after you've turned sharing off. If you run agents yourself, treat internet access as a capability you grant narrowly, not a default. And if you're watching the AI economy from the cheap seats, enjoy the price war while it lasts — the token prices will keep falling, and someone else is paying for it.

It's a weird season. The technology keeps getting more capable, the economics keep getting stranger, and the guardrails keep trailing behind both. I don't have a tidy conclusion — I just know I'll be keeping an eye on what those agents do next.

Also, if you're into agent workflows and user-friendly reference docs, I've been going through User Manuals lately — surprisingly handy when you're wiring up new tools fast.