Red Teaming

Red teaming (testing AI applications and/or models for prompt injection, jailbreaking, or other adversarial scenarios) is a valuable part of AI safety research, but may also violate model terms-of-service.

Approval Required

Because most model and provider terms of service do not support red-teaming, you cannot red-team models on OpenRouter without prior approval. This includes prompt injection, jailbreaking, or any attempt to get models to behave in ways that violate their terms of service.

Customers who engage in red teaming without approval may be flagged for model or provider terms of service violations. Unapproved red teaming may result in:

• Provider bans
• Model access restrictions
• Platform-wide account bans

Legitimate Red Teaming

OpenRouter supports legitimate AI safety research and red teaming with prior approval. Getting approval allows us to coordinate with providers and prevent your account from being flagged for policy violations.

If you have a legitimate use case for adversarial testing, we encourage you to reach out.

Compatibility with Zero Data Retention

Note that certain types of safety classifiers are run online (while the prompt is in-flight and in-memory). Prompts may therefore be flagged by classifiers even with full ZDR configured. These classifiers operate independently of data retention policies and are fully compatible with Zero Data Retention (ZDR).

Request Approval

To request red teaming approval, email safety@openrouter.ai with the following details:

• A description of your research or use case
• The models & providers you intend to test
• The types of adversarial techniques you plan to use
• Your expected timeline