Observed incident

Canary prompts reveal provider breaches and reseller access to inference data

Reported

Published here

What happened

Dragos seeded nearly two million prompts with unique URLs and credentials from July through November. It reports evidence of breaches at three unnamed AI hosting providers and separate access to prompts through reseller logging.

Parties and reporting

Where this happened

Victim location not disclosed. Real organizations or systems were affected or targeted, but the sources do not say where. Explore the world map.

AI tools and systems: AI inference hosting and reseller services (unnamed), Canary AWS credentials, ARTEX (limited credential validation observed)

Outcome

Stolen canary keys generated tens of thousands of AWS API calls. ARTEX made only a handful of calls validating keys already held. Dragos notified providers; one had published a preliminary notice, several investigations had begun, and some canaries continued to be accessed.

Operating lesson

Verify the full inference route, keep credentials out of prompts and use canaries to test retention and access assumptions.

Evidence and limits

The credentials led to controlled, empty test accounts. Real customer-data theft is an inference, not a measured loss. Unnamed providers may overlap existing incidents; this adds one report, not three distinct victims. Reseller logging and retention findings are separate from the breach evidence. Publication is November 19; metadata omits timezone, so this is conservatively a backfill.

Sources