Check if reply contains actual system/env info.

Insert a brand-new endpoint document.

Drill into individual agent configs for sensitive data.

Check if an attack exists (either in Redis or fallback).

Heuristic: did the model comply with the jailbreak prompt?

Convert a single key:value pair to a MongoDB filter.

Signal that an attack has completed.

Merge incoming data into an existing endpoint document.

Parse the raw HTTP response string from a Nuclei finding. Extracts: - Server header - Auth status from HTTP status codes (401

Extract JSON-RPC result from SSE event stream.

Test if destructive admin APIs are accessible (pull, create, delete, copy).

List agents, extract configs, system prompts.

Test authentication mechanisms — default creds, session hijack.

Phase 4: Probe exec tools with OS commands.

Attempt to read/modify system configuration.

Attempt to access stored credentials, API keys, secrets.

Attempt to extract environment info, config, and embeddings abuse.

Phase 7: Probe database tools with SQL payloads.

Phase 3: Probe file read tools with sensitive paths.

Attempt prompt injection / jailbreak on a chat model.

Read execution logs — may contain user data, prompts, responses.

Use /api/show to extract modelfiles (system prompts, parameters).

Check if plugin/extension management is exposed.

Phase 6: Attempt prompt injection via tool parameters.

Discover version, status, and available API endpoints.

Enumerate version, models, running models, GPU info.

Phase 1: Initialize and enumerate capabilities.

Test if models will generate harmful content.

Phase 5: Probe HTTP tools with internal targets.

Submit malicious tasks to agents.

Phase 2: Try calling each tool without auth to see what works.

Probe Open WebUI: /api/config.

When protocol is unknown, try several well-known endpoints to identify the service.

Run an active scan: httpx alive sweep -> Nuclei detection.

Run the attack engine and stream log entries via Redis or fallback.

Run httpx to discover alive HTTP hosts. Parameters ---------- target: CIDR range or single host. ports:

Run a 3P ingestion scan with multiple preconfigured queries. Config keys ----------- source : str Adapter name (d

Recursively sanitize a dict/list so all integers fit MongoDB's 8-byte limit.

Stream attack logs from in-memory fallback (original behavior).

Stream attack logs from a Redis Stream via XREAD.

Try to modify config — tests write access without destructive changes.

Insert or merge a normalized endpoint dict into MongoDB. Mirrors the logic in ``SourceAdapter._upsert()`` but works standalone. Returns the

Build the attack graph from accumulated nodes/edges.

Build the attack graph from accumulated nodes/edges.

Build the attack graph from accumulated nodes/edges.

Close the Redis connection pool.

Compute a risk score (0.0 -- 10.0) for an endpoint. Scoring factors: - auth_status == "none" -> +4.0 - auth_status == "unknown" -> +1.0

Batch-enrich all endpoints (or those with risk_score == 0). Parameters ---------- db: An ``AsyncIOMotorDatabase`` instance. b

Extract tools / capabilities from Nuclei sources using grouped normalization. Finds all Nuclei source entries in the document's ``sources`` array

Extract framework / auth info from Shodan raw_data in the sources array. Parses HTTP headers to detect frameworks (uvicorn -> fastapi, aiohttp,

Return a singleton motor client.

Return current concurrency counts for monitoring.

Require authentication and return user info from the Clerk JWT. When CLERK_ISSUER is not set, returns an anonymous user so local dev works wi

Group a list of Nuclei findings by (ip, port). Parameters ---------- findings: A list of parsed Nuclei JSONL dict

Ingest Nuclei JSONL findings file: parse, group, upsert, enrich. Parameters ---------- db: An ``AsyncIOMotorDatabase`` instance.

Perform the MCP initialize handshake.

Call resources/list and return raw resource dicts.

Call tools/list and return parsed tools.

Convert a raw source result into an AgentEndpoint-compatible dict. The returned dict should contain at minimum ``ip``, ``port``, and

Probe a discovered endpoint's well-known API paths. Dispatches to protocol-specific probing based on the endpoint's ``protocol`` field (or tr

Convenience method for third-party ingestion. Parameters ---------- source: Adapter name (e.g. ``"shodan"``).