For most evals you don't need paragraphs of generated text. You need a calibrated choice: is this grounded? is this a jailbreak? which answer is better? That's a classification, not an essay.

Most AI agent "security tests" are one-off prompts in a notebook. agent-probe runs 12 attack probes mapped to the OWASP Top 10 for LLM apps, outputs SARIF, and fails your CI when the agent gets…