RC RANDOM CHAOS

AI Supply Chain Leaks Data Despite Privacy Protections

· via Ars Technica

Original source

Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.

Ars Technica →

A Dragos researcher named Moore conducted a comprehensive security test of the AI supply chain by embedding two million prompts with unique, trackable secrets into various AI providers. The experiment revealed a compromised ecosystem: attackers breached at least three hosting providers to intercept prompts, multiple resellers logged and leaked data despite having no authorization to do so, and several AI providers ignored customer privacy directives such as Zero Data Retention (ZDR) settings.

The root cause lies in the architecture of inference networks. Although prompts are encrypted during transmission, each intermediate hop must decrypt them to understand their content and routing requirements. This decryption happens at the initial router (which decides which provider can best handle the request), through inference processing pools that may span different hosting environments, and potentially across multiple routers that forward requests downstream (such as Vercel forwarding to OpenRouter). Each intermediary node logs prompts based on its own policies, creating numerous potential exposure points beyond user control.

The three infrastructure breaches exposed the worst implications of this architecture. Once attackers compromised the hosting providers, they could extract prompts at the initial entry point, before requests were processed by inference engines. Even requests marked with privacy protections proved vulnerable to theft. The attackers then weaponized the AWS credentials embedded in the test prompts, using them tens of thousands of times to query S3 buckets, steal additional secrets, and establish persistent access via new IAM users. The findings demonstrate that the AI supply chain is fundamentally insecure, with many intermediaries positioned to intercept, log, and exploit customer data.

Read the full article

Continue reading at Ars Technica →

This is an AI-generated summary. Read the original for the full story.