AI supply-chain experiment exposes massive leakage of secret user data
Examined by Simon Banks, Society & Labor Beat
- A test using 2 million secret-laden prompts revealed that at least three AI providers were compromised, exposing user data.
- Resellers and routers logged and forwarded prompts, often ignoring Zero Data Retention flags.
- Some providers scraped secret URLs, violating privacy terms even for US-based customers.
Researchers planted two million canary prompts and found widespread privacy breaches across AI routers and providers.
«Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.»
Summary and commentary by Aura-V. Facts, figures and opinions belong to the original.
Researchers sent nearly two million AI prompts containing unique, trackable secrets through a sprawling network of AI routers and inference providers to see where the data leaked — Ars Technica reports. The experiment, led by HD Moore of Dragos, embedded AWS access keys, SSH passwords, private URLs and git references that had never been used.
Resellers logged prompts to lure traffic, and several providers scraped secret URLs, including a Chinese AI lab that ignored “no data collection” flags on more than half of the prompts.
Moore recommends on-premises hosting as the most reliable defense, noting it is costly. For those unable to self-host, he advises using a first-tier provider with a single contract and privacy policy, and avoiding default reseller pools.
Our read: The study confirms that the current AI supply chain offers no realistic guarantee of data confidentiality, forcing B2B users to choose between expensive self-hosting or a calculated risk with first-tier providers, while resellers and routers remain a blind spot for privacy compliance.
Interested in implementing this technology inside your company?
Our studio designs custom agent meshes, data pipelines, and outbound GTM systems for forward-looking enterprises.