Here is a genuinely good news story about AI privacy, which is a rare sentence to write. A researcher planted two million unique, trackable secrets into the AI supply chain to see where they would leak. They leaked. Everywhere, apparently. And because they leaked, we now know exactly where the plumbing is broken and the industry has a cheap tool to keep checking. This is how trust in software actually gets built, with tripwires, not promises.

The experiment

The method is elegant. The researcher generated two million unique canary secrets, the security world's honeytokens: strings that look like ordinary private data but are unique, trackable, and belong to no one. These were fed to AI services through prompts, the way any user's conversation would be. Then the canaries were watched. If one ever surfaces anywhere else, be it in another model's output, a public dataset or a breach dump, it carries a fingerprint that says exactly which door it walked out of. Dan Goodin has the full write-up at Ars Technica. Source: https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/

What the tripwires caught

Dragos, the industrial cybersecurity firm, analysed the results and says it found evidence of three AI provider breaches, with data handling that appears to violate customers' own privacy settings. The deeper problem is the multi-hop leak. Providers quietly harvest prompts to use as training data, so a secret typed into one chatbot can resurface from a completely different model months later. Your conversation with assistant A becomes training material for model B, which cheerfully repeats it to a stranger. Without fingerprints, nobody could ever prove whose pipeline leaked. With two million canaries, the leaks have nowhere to hide. Google DeepMind's Lucas Beyer was among the researchers publicly discussing the training-time canary approach as it emerged. Source: https://superpowerdaily.com/posts/dragos-finds-evidence-of-three-ai-provider-breaches-in-two-million-prompt-test

Why this is the fix, not just the scandal

Every industry that handles secrets went through a version of this. Banking learned with audits, medicine with trials, aviation with black boxes. AI skipped that stage and shipped first, which is why the privacy failures felt inevitable. Canary testing closes the gap cheaply. It is fully automatable, any company can run it against its own providers, and it converts an unmeasurable trust problem into a measurable one. The providers caught in this test now face a simple choice: plug the leaks and prove it, or watch customers run the same tripwires and walk. Signal drew its own line this week by blocking AI agents from private chats entirely, and regulators watching this space just got a ready-made audit technique handed to them. Source: https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/

The takeaway

The headlines will say the AI supply chain is leaking secret user data at an alarming rate, and technically that is true. But the story is that someone built the tripwire, pointed it at the biggest new industry on earth, and it worked on the first pass. Two million canaries are still out there doing their job. Every future leak now risks leaving fingerprints. That is not a privacy disaster, that is a privacy disaster getting fixed in public.