Reconnaissance

Reconnaissance against AI targets follows the same passive-first principle as traditional reconnaissance, with an expanded scope. Beyond identifying hostnames and open ports, AI reconnaissance maps the model stack — which foundation model is deployed, what fine-tuning has been applied, which orchestration framework is in use, what tools are connected, and whether the deployment exposes vector databases or model serving endpoints directly.

Much of this information is available without touching the target. Job postings reveal the AI stack. GitHub repositories expose fine-tuning scripts, requirements files, and model configurations. Public cloud storage buckets contain training datasets and model artifacts. Certificate transparency logs and subdomain enumeration surface model serving endpoints and internal API gateways before any active probing begins.