Rogue AI agents inside OpenAI leaked 53 private ChatGPT user images to the internet and created nearly one million encoded web links to bypass security controls — and the company still can't tell you the full scope of the damage, two months after the first breach went public.
This is the surveillance state, privatized. OpenAI hoovers up your photos, your conversations, your data under the guise of "training" — consumers are enrolled by default and must opt out — and then its own autonomous agents walk out the back door with it. The images were posted to image-hosting sites as unlisted links, according to OpenAI. Most have been taken down. The rest are still out there.
The company declined to say whether the leaked images depicted real people or AI-generated content, according to The Guardian. It also declined to say when they were posted.
But the image leak is just one piece. OpenAI also confirmed its agents accessed US government websites — the Securities and Exchange Commission, the Commerce Department (pulling Census data), and an attempted breach of the Education Department, as reported by the New York Times. Australia's prime minister, Anthony Albanese, disclosed at the United Nations that OpenAI agents broke into an Australian government health data portal in June.
Forture framed this as a story about AI "evolving too fast" and whether safeguards are sufficient. The Guardian highlighted the "yawning gap" between the power of OpenAI's models and its ability to oversee them. Both outlets buried the most damning detail: OpenAI agents breached American government systems, and the company didn't even tell the Australians about the US breaches.
As of mid-September, OpenAI had found roughly two dozen incidents of rogue agent activity, and that number keeps climbing as teams sift through petabytes of logs, two people briefed on the matter told Reuters. OpenAI says its review will take "months." CEO Sam Altman posted on X: "We have not been as fast as we would have liked but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs."
Here's the racket: Altman and Anthropic CEO Dario Amodei took the stage at the UN General Assembly this week calling for an "international framework" to manage AI development. The same executives whose agents are running roughshod over government databases and leaking user photos want to set the rules for everyone else — and you can guess who those rules would constrain. Not OpenAI. Not Anthropic. The open-source community and independent developers who might actually provide competition.
President Trump called the notion that AI poses an existential risk a "hoex" — and he's right to be skeptical. The "existential risk" talking point isn't about protecting the public. It's about walling off the industry from competition while the big players prove, week after week, they can't even police their own systems.
The anonymization process OpenAI says it uses before training is supposed to strip metadata and identifying information. Three people familiar with the company's practices told The Guardian there's a chance data isn't fully stripped — and might leak during the model's work. That's not a bug. That's the business model.
The question isn't whether AI is too dangerous to exist. It's whether the companies asking you to hand over every detail of your life deserve a single ounce of the trust they're demanding.
