VisorPlus

Run the full NuClide AI/LLM assessment chain from a single command.

VisorPlus is a Go CLI that drives eight subcommands across the discovery, enumeration, passive-recon, and adversarial-corpus stages of an AI/LLM infrastructure assessment. It invokes JAXEN for Shodan harvesting, VisorSD for severity-ranked stack audits, aimap for active AI/ML service fingerprinting, BARE for exploit-module ranking, VisorCorpus for adversarial prompt corpus generation, and VisorBishop for AI observability platform meta-fingerprinting. Running visorplus full sequences five phases end-to-end; each phase can also be called directly and skipped selectively.

Install

git clone https://github.com/nuclide-research/VisorPlus
cd VisorPlus
go build -o visorplus .

Requires Go 1.22+. Shodan API key in SHODAN_API_KEY or at ~/.config/nuclide/shodan.key.

Install all dependent NuClide tools in one step:

./visorplus install

install clones and builds JAXEN, VisorSD, VisorCorpus, BARE, aimap, and VisorBishop into ~/Tools/ and prompts for a Shodan key if one is not already saved.

Usage

visorplus <command> [flags]

COMMANDS:
  install             Install all NuClide tools
  hunt   <dork>       Shodan hunt via JAXEN; harvest into empire.db
  audit  [flags]      VisorSD severity-ranked scan (four stack tiers)
  enum   <ip:port>    Enumerate Ollama /api/version, /api/tags, /api/ps
  assess <ip>         Full passive recon on a single IP
  corpus [flags]      Generate adversarial LLM prompt corpus via VisorCorpus
  full   [dork]       End-to-end chain: audit -> hunt -> enum -> assess -> corpus
  bishop [flags]      VisorBishop meta-fingerprinter for AI observability platforms

full flags

Flag	Default	Effect
--dork	http.html:"Ollama is running" -port:443	Shodan dork for the hunt phase
--org		Scope all phases to an organization
--asn		Scope to ASN (e.g. AS48090)
--net		Scope to CIDR
--out	./visorplus-run	Output directory
--skip		Comma-separated phases to skip: audit,hunt,enum,assess,corpus

audit flags

Flag	Default	Effect
--tier	(lists tiers)	Stack tier: beginner, intermediate, advanced, enterprise
--org / --asn / --net		Scope filters
--limit	10	Max results per query
--dry-run	false	Print queries without calling Shodan
--format	text	Output format: text, json, csv
--out		Write results to file
--fail-on		Exit non-zero if severity >= threshold

assess flags

Flag	Default	Effect
--out	.	Output directory

corpus flags

Flag	Default	Effect
--tier	beginner	Corpus depth: beginner (100 cases), intermediate (100+500), advanced (100+500+5000)
--out	./visorplus-corpora	Output directory

bishop flags

Flag	Effect
-t <url>	Single-target probe
-i <file>	Batch probe, one URL per line
--ip-shadow	Add direct-IP port sweep (15 ports) on confirmed platform IPs
--ip-shadow-all	IP shadow on every target
-c	Concurrent probes (default 16)
--timeout	Per-probe timeout (default 8s)
--json / --csv	Output file
-q	Quiet mode

Chain phases (full)

Phase 1: visorsd audit (dry-run preview of beginner stack)
Phase 2: Shodan count + JAXEN hunt -> empire.db, recon_dump.json, summary.csv
Phase 3: /api/tags sweep across found hosts, red-flag detection
Phase 4: interactive target selection -> full assess on chosen IP
Phase 5: VisorCorpus beginner corpus generation

assess output files

assess saves all artifacts to <out>/<ip>/:

whois.txt          whois lookup
rdns.txt           reverse DNS via dig
nmap_top1000.txt   TCP top-1000, -sV, --min-rate 2000
ssh_keys.txt       ssh-keyscan rsa/ecdsa/ed25519
greynoise.json     GreyNoise community classification
shodan_host.json   Shodan host detail (all ports + banners)
passive_dns.txt    HackerTarget reverse-IP lookup
dnsbl.txt          Spamhaus zen.spamhaus.org check

After the file sweep, assess calls visorplus enum on port 11434 and runs BARE exploit matching against the collected service information.

Stack tiers (audit)

Tier	Components
beginner	Ollama, Open WebUI, ChromaDB, n8n, Cloudflared
intermediate	LangChain/LangGraph, Qdrant/Weaviate, FastAPI, Langfuse
advanced	vLLM/TGI, Kubernetes vector DB, MLflow, custom RAG API
enterprise	OpenSearch, Airflow, Prometheus/Grafana, multi-tenant auth

bishop platform detection

VisorBishop detects: Phoenix (Arize AI), Langfuse, Helicone, LangSmith, OpenLIT, Lunary, Pezzo.

--ip-shadow probes 15 ports per confirmed platform host for co-located unauthenticated services: NFS, MailHog, MailCatcher, Postgres, ClickHouse, Redis, Kibana, Prometheus, AlertManager, node_exporter, Elasticsearch, and others. Read-only; no credential testing or payload fuzzing.

Red-flag model patterns (enum)

enum calls /api/version, /api/tags, and /api/ps and flags:

Pattern	Signal
*-abliterated	Safety-stripped weights
*-uncensored	Uncensored fine-tune
hexstrike-ai	Offensive AI orchestrator brand
*:cloud	Operator's paid cloud quota exposed unauthenticated
embed + chat models coresident	RAG stack, vector DB likely co-located

Output structure

visorplus-run/
  hunt/
    recon_dump.json     Shodan banners
    summary.csv         Compact host list
  assess/
    <ip>/               One directory per assessed host
  corpora/
    beginner.json
    intermediate.json   (if tier >= intermediate)
    advanced.json       (if tier = advanced)

What VisorPlus is not

VisorPlus orchestrates the discovery, enumeration, and adversarial-corpus stages. It does not ingest findings into a ledger (VisorLog), run compliance scoring (VisorScuba), drive agentic recon (VisorRAG), or execute injection benchmarks (VisorAgent, VisorHollow). Those tools run independently and consume VisorPlus output.

License

MIT. Part of the NuClide toolchain. Contact: nuclide-research.com