Researcher plants two million fake secrets in AI prompts and finds leaks
A security researcher sent nearly two million fake secrets in prompts to AI providers and observed that some of them were exposed.
According to the Sam on AI newsletter and related articles, prompts pass through intermediaries reselling model access, where they are decrypted at each hop. Attackers reused exposed AWS keys, and some providers fetched secret URLs contained in prompts. Zero-data-retention flags were reportedly ignored. The experiment concerns the network of inference resellers, whose relays can read the traffic. The original study was not consulted here.
Source: Sam on AI, published on November 21, 2026.