GET /vlai/attack-techniques/{technique_id}/vulnerabilities

Vulnerabilities whose description is nearest to one MITRE ATT&CK technique in the CIRCL bi-encoder space, best first. A similarity search over the indexed descriptions, not a classification: only the vulnerability -> technique direction has measured accuracy. in_vocabulary is false for techniques the model was not trained on, which rank noticeably worse. score is the training-time probability sigmoid(logit_scale * cosine + logit_bias).