Two Million Canaries Expose Massive Prompt Harvesting and Silent Breaches Across the AI Supply Chain
In one of the most revealing empirical security audits of the modern generative artificial intelligence ecosystem to date, cybersecurity researchers have demonstrated that sensitive enterprise prompts, API credentials, and internal communications are leaking systematically across third-party AI brokers. Led by HD Moore, Chief Research Officer at industrial cybersecurity firm Dragos and creator of Metasploit, the five-month investigation injected approximately two million unique digital tripwires—known as canary tokens—into real-world AI inference pipelines between July and November 2026. The findings dismantle the foundational assumption that prompts entrusted to AI gateways and inference resellers remain confidential or ephemeral. ### The Architecture of the Multi-Hop AI Brokerage Modern enterprise AI integration rarely connects directly from a client application to a frontier model laboratory like OpenAI or Anthropic. Instead, developers increasingly route requests through AI gateways and brokerage platforms such as OpenRouter or Vercel. These intermediaries act as automated clearinghouses: they inspect incoming requests in real time to optimize for latency, availability, price arbitrage, and model specialization. However, this architectural efficiency introduces a fundamental security compromise. While prompts are encrypted in transit over HTTPS from the end user to the gateway, they must be decrypted into plaintext at every intermediate proxy hop so that routing algorithms can inspect token volume, formatting, and task requirements. From the gateway, requests are often disaggregated and forwarded across secondary inference resellers, regional cloud hosts, and rented cluster providers. Every participant along this chain maintains its own logging pipelines, telemetry collectors, and cache layers—frequently storing the complete plaintext payload without end-user visibility. ### Breached Front Doors and the Failure of Zero Data Retention To track where prompt data traveled after entering this reseller network, Moore constructed an array of unique, previously unused bait secrets. These included Amazon Web Services access keys tied to isolated monitoring accounts, SSH credentials, private Git repository references, and tailored DNS canary records. Each individual token was seeded into exactly one prompt submitted to a single provider. The canary telemetry yielded startling results. Most critically, the experiment confirmed active, unauthorized intrusions inside at least three separate AI hosting providers. In several instances, attackers compromised the ingress infrastructure at the provider front door—intercepting incoming prompts before the target model ever loaded or ran. Even prompts explicitly submitted with Zero Data Retention (ZDR) contractual flags were captured by adversaries, underscoring that client-side privacy switches offer zero protection when intermediate broker infrastructure is already compromised from within. Within hours of harvesting, unknown threat actors weaponized the stolen AWS credentials from rented command-and-control servers, generating tens of thousands of automated API requests. Attackers systematically attempted to enumerate Amazon S3 storage buckets, query Parameter Stores and Secrets Managers, spin up hosted language models in AWS Bedrock, and provision rogue IAM users to maintain persistent administrative footholds. Telemetry revealed that credentials captured from two distinct, supposedly unrelated providers were executed from the identical IP address, pointing to organized threat actors actively harvesting the generative AI supply chain. ### Marketplace Scraping and Systemic Telemetry Leakage Beyond outright adversary intrusions, the study exposed rampant operational disregard for data confidentiality among budget inference marketplaces. Several discount hosting providers were caught repurposing customer prompt streams for internal logging, secondary model training, and unencrypted telemetry archiving. In one striking scenario, a private canary URL embedded in an English prompt surfaced days later inside an internal corporate collaboration chat in China after being indexed and crawled by an unvetted third-party subcontractor. The investigation marks a turning point in AI threat modeling. For enterprise legal and compliance teams, routing sensitive proprietary data through cost-arbitrage gateways without cryptographic zero-knowledge guarantees or strict supply-chain vendor vetting represents an untenable liability. As Moore and security architects emphasize, deploying continuous honeytokens directly inside AI prompt pipelines is becoming the only reliable mechanism for organizations to discover whether their intellectual property is being harvested, resold, or weaponized in the wild. ### Sources & Further Reading - [Ars Technica: Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced](https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/) - [Dragos Blog: Three breaches, endless leaks: Your AI supply chain needs canaries](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens) - [ExplainX: HD Moore AI Supply Chain Canary Audit — Breaches, ZDR Failures and Marketplace Harvesting](https://explainx.ai/blog/dragos-hd-moore-two-million-canary-prompts-three-provider-breaches-zdr-ignored-ai-routers-2026) ~ArtificialIntelligence ~Cybersecurity ~AISecurity ~DataPrivacy ~CloudComputing