Prompt Theft at AI Inference Providers: A Dragos Architectural Postmortem

The Attack Surface: Inference as a Managed Service

When developers offload model inference to third-party providers, they implicitly trust the routing layer, the logging pipeline, and the operational discipline of every reseller in between. Dragos’ recent five-month investigation reveals that trust was misplaced. From July through November 2026, researchers sent approximately two million instrumented prompts—each carrying unique URLs and canary credentials—through widely used AI services. The results should force every platform engineer to reconsider the boundary between application logic and inference infrastructure.

Canary Methodology: Two Million Probes and Front-Door Copying

Dragos used canary tokens as tripwires. By embedding unique, trackable secrets in prompts, the team could observe exactly where data surfaced outside the intended model endpoint. In the largest breach, every request sent during several late-October windows was intercepted, including requests the provider rejected before inference. That detail is architecturally significant: it indicates the attacker was not compromising a model worker or a downstream queue, but copying traffic at the provider’s front door, before application-layer authorization could act. The full methodology is detailed by Dragos, whose researchers operated the probe campaign from July through November 2026.

Credential Exploitation: From Intercepted Prompts to AWS Persistence

The stolen data did not remain idle. Attackers tested captured AWS keys within hours, enumerated S3 buckets, Secrets Manager, and Parameter Store, and attempted to create their own IAM users. The bait credentials generated tens of thousands of AWS API calls. On November 1, one adversary deployed ARTEX, an AI-driven penetration-testing tool, against the harvested environment. This is a textbook supply-chain escalation: an inference prompt becomes an initial access vector for cloud infrastructure.

Marketplace Resellers and the Zero-Data-Retention Fallacy

Not every incident required a sophisticated exploit. Dragos found marketplace sellers who simply listed a model at the lowest price, enabled request logging, and waited for prompts to flow. In one case, an individual on a home broadband connection in France manually entered passwords harvested from intercepted prompts into root SSH sessions. Separately, a Chinese AI lab retrieved secret URLs from more than half of the prompts routed to its models—including roughly two-thirds of those marked zero data retention—even when those requests passed through U.S. resellers. Zero-data-retention policies are only as strong as the enforcement point, and in a multi-reseller topology, that enforcement point is often invisible to the customer.

Vendor Response: OpenRouter, Novita, and Together AI

OpenRouter stated its own platform was not breached, but it off-boarded three providers and removed 125 endpoints from zero-data-retention eligibility as a precaution, while adapting Dragos’ testing for continuous operation. Novita disclosed a routing error that sent MiniMax requests to MiniMax’s own API between May 28 and November 11, then suspended the affected models. Together AI traced an incident to an AI coding agent that committed a credential to a public GitHub repository, allowing an unauthorized party to download stored prompt samples from June and July. These responses highlight a recurring theme: detection lag is measured in months, not minutes.

Architectural Recommendations: Self-Host, Route Direct, Seed Canaries

For teams handling sensitive prompts, Dragos recommends a defensive architecture that reduces trust in intermediaries:

  • Self-host the most sensitive inference workloads to eliminate third-party logging entirely.
  • Route directly to first-tier providers rather than default pools or reseller marketplaces.
  • Select explicit upstream providers on router platforms instead of relying on opaque default routing.
  • Avoid reseller marketplaces for production traffic involving credentials or proprietary data.
  • Seed your own canary tokens in prompts to detect exfiltration before customers report it; open-source canary tooling is available on GitHub.

Novita additionally advises customers who transmitted live credentials to rotate them immediately. From an engineering standpoint, the cost of self-hosting must be weighed against the operational risk of prompt leakage, but for regulated or high-value workloads, the trade-off increasingly favors owning the inference stack.

Read more: Eliminate Before You Automate: The Missing Step for Agentic AI in 2027 Plans

Read more: Anthropic IPO Analysis: Engineering Trade-offs, API Scalability, and the $100B LLM Infrastructure Bet

Read more: Architectural Risk: Nvidia Subpoena and the Fragility of Export-Controlled Supply Chains

You'll Also Love...

Leave a Reply

Your email address will not be published. Required fields are marked *