Researchers recently planted two million unique identifiers into requests that travel through a network of AI routing services and demonstrated that sensitive prompt data can be accessed outside the intended model. The experiment shows that the routing layer—often an opaque broker such as OpenRouter or Vercel—does not guarantee privacy, prompting engineers to reassess how they handle user prompts.
What the experiment revealed about AI router data leakage
The study sent specially crafted prompts containing trackable URLs through popular AI routers. When the prompts reached a China‑based model provider called MiniMax, the embedded URLs were accessed, confirming that the router or one of its resellers—AtlasCloud, Novita, or StreamLake—read the data. OpenRouter later notified customers that some prompts may have been read and removed a “Zero Data Retention” label from MiniMax endpoints.
Architectural considerations for AI routing services
AI routers sit between end‑user applications and inference engines, selecting models based on price or suitability. This middle layer introduces an additional trust boundary that is not visible to the application developer. When a router forwards a prompt, the content traverses the router’s infrastructure and any downstream resellers before reaching the model. If any of those components log or otherwise process the raw prompt, the data can be exposed.
- Using a router adds a third‑party data handling step beyond the primary model provider.
- Reseller chains can amplify exposure; the same prompt may be seen by multiple entities.
- Claims such as “Zero Data Retention” may be withdrawn after evidence of leakage.
Operational and security implications
Practitioners should treat prompt data as potentially observable by any routing service they employ. This influences several operational decisions:
- Vendor evaluation: Verify the data handling statements of both the router and the downstream model provider. Look for recent disclosures or changes to retention policies.
- Prompt sanitization: Avoid embedding secrets, personally identifiable information, or proprietary code in prompts that travel through routers.
- Monitoring: Implement logging of outbound prompts and compare against known canary patterns to detect unexpected reads.
- Architecture choice: Consider direct integration with a single model provider when privacy is critical, bypassing routers entirely.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
AI router data leakage is now a documented risk. Engineers should audit any routing component in their AI pipelines, treat it as a potential data exposure point, and adjust design or operational practices accordingly. Immediate actions include reviewing router contracts for data‑handling guarantees, removing sensitive payloads from prompts, and adding canary‑style monitoring to detect future leaks.
