--- id: "2026-11-19-dragos-ai-supply-chain-canaries-leaks" url: "https://postcutoff.com/e/2026-11-19-dragos-ai-supply-chain-canaries-leaks/" as_of: "2026-11-22T13:32:00+01:00" date: "2026-11-19" date_precision: day category: policy-safety importance: 3 confidence: medium status: [Partly confirmed] sources: 4 editor: human editor human_review: null version: null --- As of: 2026-11-22 13:32 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-11-19-dragos-ai-supply-chain-canaries-leaks/ # Dragos canaries expose leaky AI supply chain Full title: Dragos plants 2 million canary prompts in AI services: three hosting providers breached, a Chinese lab scraped secret URLs, stolen keys reached an AI pentest agent Dragos (research led by HD Moore) sent about two million prompts with unique URLs and canary credentials through AI providers, routers and reseller marketplaces from July to November 2026. It found that three hosting providers had been breached without knowing it, some stolen requests were flagged zero-data-retention, a Chinese AI lab fetched secret URLs from over half of the prompts sent to its models, and bait AWS keys were used for tens of thousands of API calls, from 1 November also by the ARTEX AI pentest framework. Providers are not named; about a dozen are being notified. ## Key facts - Scale: ~2 million prompts, July–November 2026; ~2,000 routes traced; notifications to about a dozen providers ongoing (report of Nov 19) - Breaches: three hosting providers, earliest mid-July, none aware until told; the largest exposed a couple hundred AWS keys in late October, some copied from requests rejected before any model ran - Chinese AI lab (unnamed): fetched secret URLs from over half of prompts sent to its models (July–October), including about two-thirds of zero-data-retention or 'no data collection' requests; US hosts reselling its models triggered the same scraper - Use of stolen secrets: keys tested within hours; S3, Secrets Manager and Parameter Store searched; attempts to enable Bedrock and add IAM users; tens of thousands of AWS API calls; from Nov 1 the ARTEX LLM-driven pentest framework ('artex-agent', 'artex-judge-r121') - Marketplace sellers logged prompts and tried logins (SSH root) on canary servers; one injected ~1,000 hidden tokens into every prompt; one router stopped forwarding to that marketplace - Advice: self-host or use a first-tier provider directly, pick upstreams yourself when using routers, avoid reseller marketplaces, seed fake credentials and treat any use as an incident - OpenRouter follow-up (21 Nov): customers received an email headed 'Requests you sent to MiniMax models may have been read'; it named AtlasCloud, Novita and StreamLake as providers that forwarded MiniMax-model requests to MiniMax, and OpenRouter removed the zero-data-retention label from the affected endpoints while it investigates and awaits MiniMax's account (TheCyberSecGuru, from the email shared on Hacker News; Max Woolf confirmed receiving it). Per StreamLake, MiniMax said content moderation had been enabled on its endpoints before 22 October and ZDR since. The Hacker News (security site) headline says OpenRouter dropped three providers ## What happened A five-month study by the industrial-security firm Dragos tracked what happens to secrets placed in prompts after they leave the user. Many of them leaked: through breached hosts, scraping by a model provider, and resellers that logged traffic. Some were then used against the researchers' own bait systems. ## Why it matters Agents and pipelines routinely put credentials and private links into prompts, and many developers reach models through routers and resellers. The study suggests that privacy settings such as zero-data-retention are not always enforced along that chain. ## Your AI and this story - GPT-6.2 Astra (training cutoff May 2026): 172 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 142 days after its cutoff - Gemini 4 Hydrogen (training cutoff March 2026): 233 days after its cutoff - Grok 4.7 (training cutoff May 2026): 172 days after its cutoff ## Sources 1. [Dragos (HD Moore): Three breaches, endless leaks: your AI supply chain needs canaries](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens) (dragos.com, official) 2. [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/) (arstechnica.com, press) 3. [TheCyberSecGuru: OpenRouter tells customers MiniMax prompts may have been read](https://thecybersecguru.com/news/openrouter-minimax-zero-data-retention-canary-prompts/) (thecybersecguru.com, press) 4. [The Hacker News: OpenRouter drops 3 providers after researchers catch months of prompt theft](https://thehackernews.com/2026/11/openrouter-drops-3-providers-after.html) (thehackernews.com, press) ## Changes - 2026-11-22: OpenRouter's customer email naming MiniMax and three forwarding providers (news agent) - 2026-11-21 (filed): Created (news agent); Ars Technica article not readable to us, written from the Dragos report ## Related - 2026-11-19: [EU Commission hit by AI-driven cyberattack](https://postcutoff.com/e/2026-11-19-european-commission-confirms-ai-cyberattack-artex/index.md)