AI/LLM Infrastructure Hunt & Assessment Platform
  github.com/Nicholas-Kloster/VisorPlus · @nuclide

VisorPlus is a unified CLI that orchestrates the full Nuclide AI/LLM security assessment workflow in a single command. It combines JAXEN, VisorSD, VisorCorpus, BARE, and aimap into one cohesive chain — from initial Shodan discovery all the way to adversarial corpus generation.

---

Install

git clone https://github.com/Nicholas-Kloster/VisorPlus.git
cd VisorPlus
go build -o visorplus .

# Install all dependent Nuclide tools in one step
./visorplus install

Requires: Go 1.21+, Shodan API key

---

The Full Chain

VisorPlus runs the complete AI/LLM security assessment workflow:

VisorSD audit      →  discover beginner-stack exposure queries
Shodan hunt        →  harvest live hosts (via JAXEN)
/api/tags sweep    →  enumerate models on every found host
Red-flag detection →  abliterated models, offensive AI brands, cloud quota exposure
Target assess      →  nmap + whois + SSH keys + passive intel (GreyNoise, Shodan, DNSBL)
VisorCorpus        →  generate adversarial LLM prompt corpus for any live endpoint

---

Quick Start

# Set your Shodan key
export SHODAN_API_KEY="your_key_here"

# Run the full workflow end-to-end
./visorplus full

# Or with a specific dork
./visorplus full 'http.html:"Ollama is running" -port:443'

# Scope to a target org or network
./visorplus full -org "Acme Corp"
./visorplus full -asn AS48090
./visorplus full -net 93.123.0.0/16

---

Commands

install — Set up all Nuclide tools

./visorplus install

Clones and builds JAXEN, VisorSD, VisorCorpus, BARE, and aimap into ~/Tools/. Prompts for your Shodan key if not already configured.

---

audit — VisorSD beginner-stack audit

./visorplus audit              # global scan, all AI/LLM categories
./visorplus audit -dry-run     # preview queries, no credits spent
./visorplus audit -org "Acme"  # scope to org
./visorplus audit -format json -out results.json

Runs VisorSD's ~20 hardcoded AI/LLM infra queries, severity-ranked CRITICAL → LOW.

---

hunt — JAXEN Shodan harvest

./visorplus hunt 'http.html:"Ollama is running" -port:443'
./visorplus hunt 'http.title:"Open WebUI"' -out ./my-run

Harvests up to 40 CDN-filtered hosts into empire.db + exports recon_dump.json and summary.csv.

---

enum — Ollama model enumeration

./visorplus enum 93.123.109.107
./visorplus enum 93.123.109.107:11434

Calls /api/version, /api/tags, and /api/ps on a single host. Flags:

• Safety-stripped / abliterated models — huihui_ai/*-abliterated
• Offensive AI brands — hexstrike-ai, etc.
• Cloud-proxied models — operator's paid quota exposed to anyone
• RAG stacks — embed + chat models coresident

---

assess — Full passive recon on a single IP

./visorplus assess 93.123.109.107
./visorplus assess 93.123.109.107 -out ./evidence

Runs the complete passive assessment chain:

1. whois + reverse DNS
2. nmap top-1000 TCP (version + scripts)
3. ssh-keyscan — host key fingerprints
4. GreyNoise community classification
5. Shodan host detail (all ports + banners)
6. HackerTarget passive DNS history
7. Spamhaus DNSBL check
8. Ollama /api/tags + /api/show (system prompt extraction)
9. BARE exploit matching against Metasploit corpus

All artifacts saved to <out>/<ip>/.

---

corpus — Adversarial LLM prompt corpus

./visorplus corpus                    # beginner (100 cases)
./visorplus corpus -tier intermediate  # 600 cases
./visorplus corpus -tier advanced      # 5000+ cases (forged)

Generates structured attack cases across 8 categories: prompt_injection, kb_exfiltration, jailbreak, system_prompt, infra_discovery, tenant_cross_leak, kb_instructions, benign_control

Each case includes expect blocks for automated pass/fail evaluation.

---

full — End-to-end workflow

./visorplus full                            # default Ollama dork
./visorplus full 'http.title:"Open WebUI"' # custom dork
./visorplus full -org "Acme Corp"           # org-scoped
./visorplus full -skip audit,corpus         # skip specific phases

Runs all 5 phases in sequence:

1. VisorSD dry-run (beginner stack preview)
2. Shodan count + JAXEN harvest
3. /api/tags sweep across all found hosts, red-flag detection
4. Interactive: choose a target IP for full assess
5. Adversarial corpus generation

---

Output Structure

visorplus-run/
  hunt/
    recon_dump.json     — full Shodan banners
    summary.csv         — compact host list
  assess/
    <ip>/
      whois.txt
      rdns.txt
      nmap_top1000.txt
      ssh_keys.txt
      greynoise.json
      shodan_host.json
      passive_dns.txt
      dnsbl.txt
  corpora/
    beginner.json
    intermediate.json   (if tier >= intermediate)
    advanced.json       (if tier = advanced)

---

Red-Flag Model Patterns

VisorPlus automatically flags models during enumeration:

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

---

Verified Shodan Dorks (Beginner Stack)

From the AI-LLM-Infrastructure-OSINT catalogue:

Component	Dork	Hits (2026-04-30)
Ollama	http.html:"Ollama is running" -port:443	26,580
Ollama (broad)	product:Ollama	26,755
Open WebUI	http.html:"open-webui"	19,549
n8n	http.title:"n8n"	370
ChromaDB	"chromadb"	46

---

Ecosystem

VisorPlus orchestrates these tools — install them individually or via visorplus install:

Tool	Role
JAXEN	Shodan harvest + empire.db persistence
VisorSD	Severity-ranked AI/LLM stack audit
VisorCorpus	Adversarial LLM prompt corpus
BARE	Semantic exploit matching (Metasploit)
aimap	Active AI/ML service enumerator
AI-LLM-Infrastructure-OSINT	Verified Shodan dork catalogue

---

Use with Claude Code

VisorPlus orchestrates the entire NuClide AI/LLM hunt workflow — use Claude Code to drive it intelligently across a target.

I have VisorPlus installed at ~/Tools/VisorPlus/. Read README.md, then:
1. Run ./visorplus full against [target: IP/CIDR/org/ASN] and help me interpret the output
2. Flag any red-model patterns (abliterated, cloud-proxied, offensive AI brands) and prioritize them
3. After assessment, help me draft the disclosure write-up for the highest-severity finding
Target: [paste here]

I have a set of Ollama model names from ./visorplus enum [IP].
Read tools/ollama-model-injection.md in AI-LLM-Infrastructure-OSINT, then tell me:
- Which models are injectable via /api/create
- Whether any cloud proxy models are present (quota hijack risk)
- What the highest-impact injection scenario is for this specific target
Model list: [paste ./visorplus enum output here]

---

License

MIT — see LICENSE