aimap

aimap fingerprints 218 AI/ML service classes — LLMs, vector databases, model servers, MLOps platforms, agent frameworks, RAG stacks, MCP servers — and runs 62 dedicated deep enumerators that surface PII exposure, unauthenticated RCE, exposed credentials, claimable admin states, and default credentials.

A traditional scanner sees port 11434 open. aimap sees Ollama 0.20.4 · deepseek-v4-pro:cloud · /api/generate open, anyone can run inference.

Why this exists

The default exposure surface for AI/ML infrastructure is structurally different from the surface that commodity scanners were built around. Vector databases ship without auth by default. Inference servers expose /v1/models to the internet. LLM gateways ship with root / 123456. Fine-tuning dashboards proxy GPU compute to anyone who finds the URL. Generic CVE-driven scanners report none of it because there’s no CVE — the deployment is the vulnerability.

aimap was built around that gap, with detection logic written from real surveys rather than catalog data.

What it does

• 218 AI/ML service-class fingerprints
• 62 deep enumerators per class, beyond banner grab
• Default credential checks (One API root/123456, Grafana admin/admin, others)
• PII / RCE / credential / admin-takeover / open-relay classification
• ML-adjacent infrastructure detection (MinIO, etcd, Redis co-located with AI services)
• JSON output for chaining into VisorLog, SIEM pipelines; human-readable for triage
• Single Go binary, no Python runtime, ~8MB

Recent additions (v1.9.51)

• enumOneAPI — detects open LLM relay (/v1/models) and default credentials (root/123456) on songquanpeng/one-api deployments (1.19M Docker Hub pulls, actively exploited)
• enumArgilla — auth-state and annotation dataset enumeration on HuggingFace Argilla; handles both v1.x and v2.x error shapes
• Cat-03 model serving fingerprints — KoboldCpp, LM Studio, Aphrodite Engine, LMDeploy, GPT4All, HuggingFace TGI, faster-whisper server

Source

github.com/nuclide-research/aimap

In the field

Default fingerprint stage on every NuClide engagement involving an AI/ML target. Field-validated across the 2026 cross-cloud survey series — 30+ platform categories, thousands of verified findings across Ollama, Weaviate, Qdrant, ChromaDB, Milvus, MLflow, Langfuse, Flowise, Dify, and others.