On November 19, HD Moore, Chief Research Officer at security firm Dragos, published research in which his team sent about 2 million prompts containing unique URLs and decoy AWS keys to widely used AI services and routers between July and November 2026, then tracked where those canaries surfaced. They found at least three inference providers compromised by attackers (the bait AWS keys were used for tens of thousands of API calls), marketplace resellers viewing and keeping customer prompts through built-in logging, and providers ignoring privacy settings: one Chinese AI lab fetched secret URLs from more than half of the prompts sent to its models from July to October, including roughly two-thirds of requests marked Zero Data Retention. On November 21, Novita published a privacy notice confirming some MiniMax requests had been forwarded to MiniMax's official API. The same day, Together AI disclosed that a credential committed by an AI coding agent let an attacker download prompt samples from June and July, including Dragos canary credentials.

Key Takeaways

  • ✓Scale: about 2 million prompts from July to November, each carrying a unique URL, DNS name and decoy AWS key, traceable to the exact model, route and privacy setting
  • ✓Breaches: at least three inference providers were compromised, the earliest from mid-July; attackers copied requests at the front door, and bait keys drove tens of thousands of AWS API calls, later including the ARTEX AI pentest framework
  • ✓Snooping: individual marketplace resellers only had to list a model at the lowest price and turn on request logging to read customer prompts; the leak followed the seller, not the whole marketplace
  • ✓ZDR failures: one Chinese lab fetched secret URLs from more than half of prompts (about two-thirds of ZDR requests); one link appeared in a Feishu chat 24 days later
  • ✓Follow-up: Novita confirmed some MiniMax requests from May to November, including those routed via OpenRouter, were forwarded to MiniMax's official API; Together AI disclosed that a credential committed by its coding agent let an attacker download June-July prompt samples for MiniMax M3, GLM-5.2 and Kimi K2.6, with five of the tested credentials being Dragos canaries, and now defaults new accounts to ZDR
  • ✓Advice: self-host or go direct to a first-tier provider, pin upstream providers on routers, avoid reseller marketplaces, seed your own canaries in agent context and enable secret filtering
⚡

Key Decision Metrics at a Glance

TL;DR Verdict
🏆 Benchmark RankRank
No model benchmark in this story · see current rankings
Full Leaderboard→
🛠️ Ecosystem & IDERules
AI security · Prompt leakage · Zero Data Retention
Setup IDE Rules↗
🧭

Turn your technical choice into a development budget

Compare 40 dev plans & simulate token costs vs $20/mo subscriptions

🔬

In-Depth Technical Analysis

Background

Developers increasingly rely on AI routers and inference marketplaces to cut costs, but a single request can pass through several hops, and every hop sees the full prompt and response, which often include source code, internal documents and even credentials. On November 19, HD Moore, Chief Research Officer at security firm Dragos, published research on the Dragos blog and on X that tested this supply chain with about 2 million canary prompts.

How it works

The team started with Thinkst Canarytokens and then built its own service that derives a unique hostname, URL and set of credentials for every prompt (decoy AWS keys point to empty, fully logged accounts), so any access maps back to the exact prompt, model, route and privacy setting. It also built a traceroute for AI inference, using callback image URLs and nested error envelopes from slightly broken requests to identify each hop; about 2,000 routes traced so far revealed several undisclosed middle layers.

Numbers that matter

This is security research, not a model benchmark. Dragos reports at least three compromised providers, with bait keys used for tens of thousands of AWS API calls; one Chinese lab fetched secret URLs from more than half of the prompts sent to its models, including about two-thirds of ZDR requests; and one link was posted in a Feishu chat 24 days later and opened from six devices within two minutes. Dragos did not name the providers involved. Afterwards, several users posted on X an email OpenRouter sent customers saying it removed the ZDR label from some MiniMax endpoints; OpenRouter has not published a notice on its site. On November 21 (UTC), inference provider Novita published a privacy notice on X and on its blog confirming that a routing configuration issue caused some MiniMax requests between 2026-05-28 and 2026-11-11, including requests routed to Novita through OpenRouter, to be forwarded to MiniMax's official API, where their contents may have been accessed or retained by a third party; Novita suspended the affected models and is following up with MiniMax on deleting the data. The same day, Together AI CEO Vipul Ved Prakash disclosed a security event on X and on the company blog: an AI coding agent committed a credential to Together's public GitHub repository xorl that gave access to a private Hugging Face repository of prompt samples kept for software testing (drawn from June and July requests to MiniMax M3, GLM-5.2 and Kimi K2.6); an unauthorized party used it to download the samples and search them for other credentials, five of which were Dragos canary credentials. Together has made ZDR the default for new accounts and tightened its rules for AI coding agents.

What it means for developers

Keep real secrets out of prompts and agent context, and rotate any you already sent. When using a router, pin upstream providers yourself and verify that privacy settings actually change routing, and avoid cheap reseller marketplaces; send sensitive work directly to a first-tier provider or self-host. As Dragos suggests, seed a free canary token in your pipeline context and treat any use of it as an incident.

Action HubReady to adopt this in production?

Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.

Up NextExpo Launches Expo Simulators: On-Demand Cloud iOS Simulators and Android Emulators for Coding Agents Like Claude Code, Codex and Cursor, with Automatic Session Recordings as Proof, at $0.10 per Minute
Read Now→