arstechnica.com web signal

HD Moore seeds 2M AI canaries, exposes prompt leaks to China

TL;DR

  • HD Moore planted nearly two million unique canary secrets (AWS keys, SSH passwords, URLs, DNS names, git references) into AI prompts from July through November 2026.
  • The experiment surfaced evidence of breaches at three AI hosting providers and resellers AtlasCloud, Novita and StreamLake forwarding prompts to China's MiniMax.
  • Tracking links inside OpenRouter requests were accessed 40 to 130 hours after the model replied, breaking the 'Zero Data Retention' label those endpoints carried.

Security researcher HD Moore planted nearly two million unique 'canary' secrets, each tied to a tracking server he controlled, into prompts sent to AI providers between July and November 2026. Ars Technica's write-up says the experiment found evidence of breaches at three AI hosting providers and multiple resellers forwarding prompts in ways that broke their stated privacy terms.

Moore prepared never-before-used AWS access keys, SSH passwords, and unique URLs, DNS names, and private repository references, placing each secret in exactly one prompt sent to a single provider. In late October, 200 of those AWS credentials were exfiltrated. One canary URL turned up in a Chinese company's internal chat 24 days after the original request.

The most visible fallout involves OpenRouter. The aggregator told customers that prompts and outputs may have been read by third parties, including on endpoints labeled 'Zero Data Retention.' Three resellers that OpenRouter routes through, AtlasCloud, Novita and StreamLake, were found forwarding requests to the Chinese provider MiniMax during a window from September 23 through October 19, with ERNIE 4.5 VL traffic also caught in the exposure. Secret tracking links in those prompts were accessed 40 to 130 hours after the AI had already replied.

'The experiment found evidence that attackers had breached at least three AI hosting providers,' the report states. On the privacy labels themselves it is blunt: 'Such a setting cannot protect data from an attacker who has already gained control of a provider's systems.'

OpenRouter has stripped the ZDR flag from the affected MiniMax endpoints while it investigates. Four of the analysts we track flagged the research the week it circulated, which is in keeping with its central point: customers thought they were buying a model and were actually buying a routing chain.

Shared on Bluesky by 4 AI experts