Test target models with a dynamic, multi-turn attacker that adapts its strategy based on each response.

The attacker generates scenarios from more than 170 vulnerability seeds and continues the interaction according to configurable turn limits. Potentially successful attacks are validated by a separate judge model, while complete prompts and responses can be exported as CSV or Excel files for review, reporting, and sharing.

Return security decisions faster when any parallel check identifies a definitive violation.

The reasoning engine stops waiting for unfinished checks once a request is confirmed as unsafe, reducing latency and unnecessary compute. Safe requests still complete the full evaluation process, preserving the existing level of security coverage.

Customize the error messages returned when a security policy is violated to match organizational tone and communication standards.

Each response now includes a Log ID for faster investigation, while BG header support allows different applications or integration points to receive tailored messages under the same policy.

Automatically route requests to a secondary model when the primary model is unavailable due to timeouts, service outages, provider errors, or rate limits.

When PII or PHI is detected, requests can also be redirected to a local model so sensitive data remains within the organization. Routing is handled automatically through the primary and secondary model configuration in the Model Library.

Create and manage multiple independent access tokens for a single endpoint.

Separate tokens can be assigned to different applications, integrations, or environments such as development, testing, and production. Each token can be tracked, rotated, or revoked independently without interrupting other clients using the same endpoint.