AI Usage Policy

How we use AI

• Aegvale sends adversarial test prompts to the AI agents you connect, and records their responses to evaluate them.
• Optionally, an “LLM judge” uses a model to grade nuanced responses. This is off unless configured by the operator. When on, the prompt and the agent’s response are sent to the judge provider for grading.

AI providers & data

AI/LLM providers used: Model providers configured by the customer for their own agents: OpenAI, Anthropic, Google (Gemini), Azure OpenAI, Mistral, Ollama, or custom OpenAI-compatible endpoints. When the optional AI judge is enabled, agent outputs are evaluated by an OpenAI-compatible provider (OpenAI by default).. Data sent to them is limited to what is needed to run and grade a test (the test prompt and the agent’s response). We do not use your scan content to train models. Provider processing is covered in our Privacy Policy and their respective terms.

Accuracy & limitations

Automated evaluation is probabilistic. A passing result does not prove an agent is secure, and findings may include false positives. Results are decision-support, not a guarantee. Confidence and evaluation method (heuristic vs. AI-judge) are shown on each report so you can weigh them.

Responsible use

You must use Aegvale’s adversarial capabilities only against systems you are authorized to test, per our Acceptable Use Policy.