IA · 21 November 2026 · 5 min read
Two Million Canaries Expose Massive Leaks Across the AI Supply Chain: Stolen Prompts and Compromised Secrets
In brief: A comprehensive security investigation conducted by researcher HD Moore has revealed that the widespread AI inference supply chain leaks highly sensitive data and credentials at an alarming rate. By transmitting roughly two million prompts containing traceable canary tokens through various intermediary brokers and inference providers, the study uncovered active provider compromises, prompt harvesting, and systemic failures to enforce zero-retention policies.
by Team Mocchi's
For months, software development teams have relied on inference routers and specialized gateways to reduce latency and cut costs when querying frontier machine learning models. Yet behind the appeal of affordable and abstracted infrastructure lies an opaque supply chain where confidential prompts traverse multiple intermediaries in plaintext, often rendering promised privacy guarantees completely ineffective.
The scale of this exposure was brought to light by security researcher HD Moore, chief research officer at Dragos, with technical findings detailed in a comprehensive report by Ars Technica. Moore devised an industrial-scale experiment to track whether and where sensitive information embedded in production-style prompts was being retained, siphoned, or exploited.
Two million digital tripwires across the ecosystem
To trace data spillage across the distribution chain, the researcher crafted nearly two million distinct prompts. Each prompt included a unique, never-before-used credential: fresh Amazon Web Services (AWS) access keys tied to empty honeypot accounts, SSH logins, private Git repository references, or custom DNS domain records controlled by his monitoring infrastructure. Known in cybersecurity as canary tokens, these elements were designed to trigger an alert the moment an external entity attempted to query, resolve, or authenticate with them.
The prompts were then dispatched across a wide array of intermediary services, commonly referred to as AI routers or gateways — including routing platforms such as OpenRouter and Vercel — whose core proposition is to dynamically forward client requests to whichever model provider currently offers the lowest price or best availability. Every canary was strictly mapped to an individual provider and routing configuration.
Hop-by-hop decryption and compromised providers
The resulting telemetry revealed fundamental structural weaknesses within the multi-tier routing architecture. While client-to-server traffic is encrypted in transit over HTTPS, each intermediary node must decrypt the payload into plaintext to evaluate its characteristics, calculate token limits, and choose downstream endpoints. Prompts frequently hop through several brokers before reaching physical accelerator clusters.
Throughout this journey, verbose system logs capture full request payloads. By analyzing incoming connection attempts against his canary endpoints, Moore confirmed that at least three downstream inference resellers had suffered active breaches, allowing unauthorized third parties to intercept incoming prompt streams. Furthermore, AWS keys and SSH credentials seeded inside test prompts were observed attempting logins against honeypot servers within hours of transmission, demonstrating that sensitive payload contents regularly leak into the wild.
The failure of Zero Data Retention
Equally troubling was the breakdown of contractual privacy commitments. The research documented numerous instances where prompts explicitly tagged with Zero Data Retention (ZDR) headers — flags intended to mandate immediate payload deletion post-inference — were nevertheless logged, stored, and subsequently leaked. The emergence of resellers operating atop other resellers has created a fragmented environment where developers have no visibility into the ultimate hardware operator executing their code.
Mocchi's take
This investigation dismantles the assumption that low-cost inference gateways can serve as drop-in, enterprise-grade replacements for verified enterprise endpoints. For software engineering teams building production applications that touch proprietary source code or confidential business data, routing traffic through unvetted brokerage layers introduces catastrophic supply-chain exposure. The correct engineering standard demands direct, audited agreements with verified cloud providers, combined with rigorous local sanitization and data-masking pipelines that neutralize sensitive tokens before any payload departs the corporate perimeter.