Canary tokens show AI inference providers and resellers leaking customer prompts
- Dragos sent nearly two million prompts with canary tokens to widely used AI services from July to November 2026 and saw activity suggesting data from at least three AI providers was exposed.
- Inference marketplace resellers could view and retain customer prompts through standard logging, without exploiting a vulnerability or bypassing platform controls.
- Prompts sent to U.S.-based providers were transmitted to Chinese AI organizations, and one seeded link ended up in an internal company chat in China.
- A unique URL was accessed months after a single inference request, and the login attempts against Dragos's canary servers came from home internet connections and personal computers.
Hacker News 의견들
Openrouter's zero data retention thing is basically bullshit imo. Got the half-assed breach notice and I'm not a fan.
Got that notice too, but after reading it and the together.ai one, my prompt was probably not leaked because I had ZDR turned on. Still a huge disappointment for the whole ecosystem though.
Openrouter's response says several US hosting providers resold the lab's models and prompts still flowed back to the lab. I don't see how they could offer ZDR without the providers having real skin in the game. Amazon and Google have lawsuits and backlash riding on it, these resellers don't. The trust chain is broken.
The bit about it being personal made me laugh. Home internet and personal PCs? Residential proxies have been standard in AI circles for ages, so that's not surprising at all.
Great research but the opinion section is bland. The fixes are host it yourself, go direct to a first-tier provider, or use routers with care. Self-hosting is overkill and 'be careful' is a slap on the wrist. Just fire any API token aggregator and connect to one vendor directly.