The method: traceable decoys

An experiment conducted by the chief research officer at Dragos, HD Moore, and reported by the technology outlet Ars Technica used nearly two million requests sent to several artificial intelligence service providers. Traceable test credentials had been embedded in the requests; they were created specifically for the purpose and were not taken from real users. The aim was to determine whether—and from which point in the service chain—that data would later resurface, for example in attempts to access cloud systems.

According to reports, at least three providers showed signs of compromise: the decoy keys were reportedly used for tens of thousands of calls to cloud service interfaces in an attempt to obtain other credentials, activate services, and maintain access. In some cases, the use occurred within hours of submission; in others, the credentials were stolen even when the request had been rejected before the model was run.

The chain of intermediaries and the limits of settings

A second area of risk concerns the architecture through which many AI applications distribute requests: content can pass through multiple platforms and be decrypted at intermediate steps. The research found that some resellers record and transmit requests to third parties. In addition, private addresses contained in requests labeled as zero-retention or with data collection disabled were reportedly retrieved days later, suggesting that privacy statements do not always match actual behavior across the entire supply chain.

Moore offers two recommendations: for sensitive activities, consider local deployment; for those who cannot manage their own infrastructure, go directly to providers considered trustworthy and explicitly name the upstream provider. These are recommendations from the study’s author, not measures verified by third parties.

What has been established and what remains to be clarified

The picture that emerges is that of a supply chain in which a single service setting is not enough to guarantee confidentiality, because data can be copied or intercepted before and after the model. A distinction must be drawn, however, between what is documented and what is not: the available material does not identify the three providers involved, does not quantify the exact number of compromised credentials, and does not clarify whether the resellers violated contractual obligations.

The research also does not measure the impact on real users, because the decoys were artificial. The identity of the parties involved, the actual scope of the leaks, and the effectiveness of the recommendations therefore remain to be verified. The strongest point is methodological: testing the chain with traceable markers makes it possible to identify exposure points that privacy statements alone do not reveal.

Source and attribution: this original summary was prepared by AICronaca based on the report by IT之家 ↗The statements are attributed to the source; possible information limitations: The material does not name the three compromised providers, does not indicate how many resellers recorded or transmitted the requests, and does not specify the exact number of credentials used. The conclusions come from a single study, reported by Ars Technica and amplified by IT之家, and no independent confirmations appear to exist.