Network Discovery

The discover network command fingerprints AI services using HTTP-based probes, then optionally runs vulnerability templates against discoveries.

How Fingerprinting Works

For each host:port combination:

1. TCP connect -- verify the port is open
2. HTTP probes -- send HTTP requests matching known AI service patterns
3. HTTPS fallback -- ports 443, 8443, 9443 prefer HTTPS; others try HTTP first
4. Multi-result -- all services above the minimum specificity threshold (30) are returned. Port-affinity ordering runs port-matched probes first. A specificity-100 match short-circuits the remaining probes.

Each service has one or more HTTPProbe definitions with:

• HTTP method and path
• Expected status code and/or body content
• MatchBodyNot -- reject if body contains this substring (reduces false positives)
• MatchHeader -- match response header values (case-insensitive)
• VersionRegex -- extract version from response body (first capture group)
• Specificity score (1-100) -- higher means more confident

When multiple services match on the same host:port, all are returned. When 3+ distinct services match on a single port, results are flagged as proxy_likely (likely behind a reverse proxy).

Detected Services

aipostex detects more than two dozen AI services out of the box:

Service	Default Port	Probes	Specificity	Version
Ollama	11434	GET / (body: "Ollama is running"), GET /api/tags	Medium-High	--
ChromaDB	8000	GET /api/v2/heartbeat, GET /api/v1/heartbeat	High	--
vLLM	8000	GET /v1/models, GET /health	Medium	--
LiteLLM	4000	GET /health, GET /v1/models	Medium	--
LM Studio	1234	GET /v1/models	Medium	--
LocalAI	8080	GET /v1/models, GET /readyz	Medium	--
Weaviate	8080	GET /v1/meta, GET /v1/schema	High	Extracted via regex
Qdrant	6333	GET /collections, GET /	High	--
Jupyter	8888	GET /api, GET /	Medium-High	--
MLflow	5000	GET /api/2.0/mlflow/experiments/search, GET /api/2.0/mlflow/registered-models/search, GET / (body: "OK")	High	--
Gradio	7860	GET /info, GET /	Medium-High	Extracted via regex
Streamlit	8501	GET /_stcore/health	High	--
Ray	8265	GET /api/version	High	Extracted via regex
Open WebUI	3000	GET /	Low	--
MCP SSE	3000	POST /message (JSON-RPC initialize)	High	--
MCP Inspector	6274	GET /, GET /api/servers, GET /api/tools	High	--
MCPJam Inspector	6274	GET /, GET /api/servers	High	--
OpenAI-Compatible	8000	GET /v1/models	Low (generic)	--

Note

When multiple services share ports (e.g., vLLM and ChromaDB on 8000, or LocalAI and Weaviate on 8080), the service-specific probes with higher specificity disambiguate.

Default Port List

When --ports is not specified, these ports are probed:

   80   443  8443  11434  8000  4000  8080  1234  6333
 8888  7860  8501   5000  8265  3000  6274

Ports 80, 443, and 8443 are included to detect AI services behind reverse proxies. When 3+ distinct services match on a single port, results are flagged proxy_likely and an explicit warning is emitted during the scan to alert the operator.

CIDR Expansion

The --target flag accepts CIDR notation:

./aipostex discover network --target 10.0.0.0/24     # 254 hosts
./aipostex discover network --target 10.0.0.0/16     # 65534 hosts
./aipostex discover network --target 10.0.0.0/8      # too many without --max-hosts 0

The --max-hosts guardrail (default 65536) prevents accidentally scanning huge ranges. Set to 0 to disable. Large IPv6 CIDRs (63+ host bits) are rejected before expansion to prevent integer overflow. Overlapping CIDRs are deduplicated before the host count is validated.

URL-format targets (e.g., http://10.0.0.1:8000) are normalized before CIDR expansion -- the scheme is stripped, the host is used for host processing, and any port in the URL is merged into the port scan list.

Individual hosts and multiple targets are also supported:

./aipostex discover network --target 10.0.0.5
./aipostex discover network --target 10.0.0.0/24 --target 192.168.1.0/24

Auto-Scan

By default, discover network runs vulnerability templates against discovered services (--auto-scan=true). This:

1. Maps discovered services to relevant template tags (see table below)
2. Loads and filters templates
3. Executes matching templates against each discovered service URL
4. Deduplicates results

When --tags is provided, user-specified tags replace the auto-detected service tags entirely. Only templates matching the user-provided tags will run, regardless of which services were discovered.

To fingerprint only (no template scanning):

./aipostex discover network --target 10.0.0.0/24 --auto-scan=false

Service-to-Tag Mapping

Discovered services are mapped to template tags for auto-scan filtering:

Detected Service	Template Tags
ollama	ollama
chromadb	chromadb, vectordb
weaviate	weaviate, vectordb
qdrant	qdrant, vectordb
jupyter	jupyter
mcp-sse	mcp
mcp-inspector	mcp
vllm	vllm, openai
litellm	openai
localai	openai
lmstudio	openai
openai-compatible	openai
ray	ray
mlflow	mlflow
gradio	gradio

Deduplication

Fingerprint results are deduplicated per URL+service, with all services per port preserved. Multi-result scanning means each host:port can return multiple detected services.

Vulnerability findings from auto-scan are further deduplicated using the standard finding deduplication (by source, template ID, target, title, severity).

Workflow Output

For each discovered service, discover network generates workflow recommendations grouped by host:port. Read-only commands appear first, followed by gated commands:

── Next Actions ────────────────────────────────────────────

  10.0.0.5:11434
    aipostex ollama --target http://10.0.0.5:11434 enum
    aipostex ollama --target http://10.0.0.5:11434 prompts
    aipostex ollama --target http://10.0.0.5:11434 show --model <model>

  10.0.0.6:8000
    aipostex vectordb --target http://10.0.0.6:8000 --type chromadb enum
    aipostex vectordb --target http://10.0.0.6:8000 --type chromadb search-sensitive

  Gated (require --force-exploit):
    aipostex ollama --target http://10.0.0.5:11434 poison ...

In console mode, findings are also grouped by host with detected service names in each host header, sorted by severity (critical first). Fingerprint INFO findings are folded into the host header rather than displayed individually.

These recommendations are also preserved in metadata.workflow on the corresponding findings in JSON/JSONL output.