Canary Token Test Exposes AI Provider Breaches
Canary Token Test Exposes AI Provider Breaches
Dragos seeded roughly two million prompts with unique canary tokens across widely used AI services from July through November 2026, finding evidence of three provider compromises, marketplace resellers viewing and retaining customer prompts through standard logging, and failures to honor Zero Data Retention and other privacy settings. In one case, a Chinese AI lab's systems fetched secret URLs from more than half of prompts sent to its models, including about two-thirds of requests marked "no data collection" or ZDR, and a secret link was posted into a Feishu chat 24 days after a ZDR-flagged request. Dragos did not name the providers, the breaches have not been independently confirmed, and Dragos sells security products including runZero.
The piece offers a detailed, methodologically specific account of how prompt data can leak across the AI supply chain, though its findings are vendor-reported and unconfirmed.
Source: Hacker News AI · epyhia.com