RedAmon
AI-Powered Offensive Security Platform
······ ··· ······ ······ ··· ····· ····
A parallelized recon engine maps every domain, port, endpoint, secret, and CVE into a graph database. A fireteam of autonomous AI specialists then queries that graph in real time to plan, exploit, and pivot with full battlefield awareness.
What attackers run against you, you run first.
A full offensive stack of recon engines, exploit frameworks, and AI agents, sandboxed in containers and gated by engagement phase.
One command to deploy.
Install in Three Commands
Self-hosted Docker stack. One script bootstraps Postgres, Neo4j, the agent, the recon containers and the webapp. No external accounts required to run a scan.
Your AI Offensive Engineer
An autonomous agent that thinks, plans, and attacks like a human pentester. It chooses targets, picks tools, adapts to findings, and keeps going until the objective is reached or escalation is needed.
Agent Tool Belt
The dedicated tools the agent calls directly, each with structured I/O, timeout discipline and graph-aware routing.
- query_graphNatural-language queries over the Neo4j attack-surface graph (Domains, Subdomains, IPs, Ports, Endpoints, CVEs, Secrets, Technologies). Primary recon source.
- web_searchKB-first hybrid search (FAISS + Neo4j) over GTFOBins, LOLBAS, NVD, ExploitDB, OWASP, Nuclei templates and tool docs. Falls back to Tavily when confidence is low.
- cve_intelProjectDiscovery vulnx CLI: NVD + CISA KEV + EPSS + HackerOne + public PoC links + Nuclei template availability in one passive query.
- shodanInternet-wide OSINT: search exposed devices, fetch host details, reverse DNS, domain DNS, count without burning credits.
- google_dorkPassive Google search via SerpAPI with site:, inurl:, intitle:, filetype:, intext: operators.
- execute_nuclei8000+ YAML templates for CVE verification and exploitation. Custom templates served from /opt/nuclei-templates.
- execute_curlHTTP requests for reachability, headers, status codes, banners and baseline diffing.
- execute_httpxFast HTTP prober with status, title, server, redirect chain and technology detection in JSONL output.
- execute_naabuFast TCP/SYN port scanning to verify open ports or scan targets not yet in the graph.
- execute_nmapDeep scanning with version detection, OS fingerprint and the full NSE script library.
- execute_subfinderPassive subdomain enumeration via certificate transparency, DNS datasets and search engines. No traffic to target.
- execute_amassOWASP Amass subdomain discovery with passive, active, brute-force and ASN intel modes.
- execute_gauPassive URL discovery from Wayback Machine, Common Crawl, AlienVault OTX and URLScan.
- execute_katanaJS-aware web crawler that finds endpoints, paths and known files (robots.txt, sitemap.xml).
- execute_jsluiceStatic analysis of downloaded JavaScript for hidden API endpoints, parameters and embedded secrets.
- execute_arjunBrute-forces ~25,000 parameter names to surface hidden GET, POST, JSON and XML params before injection testing.
- execute_ffufFast web fuzzer for directory, virtual host and parameter discovery with SecLists wordlists.
- execute_wpscanWordPress scanner for vulnerable plugins, themes, users and config backups.
- execute_hydraTHC Hydra brute-force across 50+ protocols (SSH, FTP, RDP, SMB, MySQL, PostgreSQL, HTTP forms, Tomcat, WordPress).
- metasploit_consolePersistent msfconsole singleton with module context and live Meterpreter sessions across calls.
- msf_restartFull Metasploit reset: kills every active session and clears module config. Takes 60-120s.
- execute_playwrightReal-browser automation for SPAs, dynamic pages, form filling, XSS dialog proof and multi-step flows.
- execute_codeRun multi-line Python, bash, ruby, perl, C or C++ exploit scripts without shell escaping. 120s timeout.
- tradecraft_lookupAuto-fetch curated security URLs (HackTricks, PayloadsAllTheThings, custom user library) and embed them into reasoning steps.
- kali_shellGeneral bash -c gateway to the full Kali toolset. See the catalog below for everything reachable from here.
kali_shell catalog
Single bash -c gateway to the full Kali toolset, wordlists, Python libraries and pre-staged post-exploit binaries.
Fireteam Parallel Specialists
Deploy N specialist sub-agents investigating independent attack angles concurrently. Each runs its own ReAct loop and approval queue.
Scatter-Gather ReAct pattern: asyncio fan-out, fan-in to unified context, no recursion, per-agent attribution.
EvoGraph Attack Chain Evolution
Every tool, finding, exploit and dead-end persists as a queryable graph that survives across sessions. New sessions auto-load prior context.
5-node attack taxonomy, dual-layer in-memory + async Neo4j writes, cross-session learning at session init.
Wave Runner Parallel Tools
When the agent spots independent tools, it runs them all at once under a single Wave card and merges their analysis.
asyncio batched tool execution, single streaming card, joint output reasoning, slashing wall-clock time by N.
Auto-Classify, Auto-Route Skills
The agent classifies every finding against the active skill registry and routes it to the matching built-in or user-defined attack skill without manual selection.
LLM tags each finding with the closest registered skill slug; phase gate, ROE check, and payload defaults are looked up automatically from the skill's manifest.
Deep Think Strategic Reasoning
At session start, phase transitions and 3+ failure loops the agent pauses for structured Situation / Vectors / Approach / Priority / Risks analysis.
Four trigger conditions, five-section output, injected into all subsequent reasoning. Counts as roughly 1 extra LLM call per engagement.
Stateful Meterpreter Sessions
Keep post-exploit shells open and run lateral movement, privilege escalation and exfil interactively, not as one-shot payloads.
Stateful + stateless modes, persistent session state, 70+ post-exploit commands, Meterpreter-aware routing.
NAT-Transparent Reverse Shells
Auto-tunnel reverse shells over ngrok or Chisel. LHOST and LPORT are auto-discovered and payloads are pre-configured for the agent.
Two providers (ngrok TCP single-port, Chisel multi-port bidirectional), auto-started inside Kali sandbox, hidden endpoint config.
10-Layer Safety Guardrails
Multi-level scope enforcement: LLM ethics gate, 250+ hard-coded .gov / .mil / critical-infra blocklist, RoE parser, stealth mode, phase gating, HITL approvals, tenant-isolated Cypher, sandboxed Kali container.
Soft LLM guardrail, deterministic regex blocklist, three-tier RoE enforcement, stealth-mode tool restrictions, phase-gated tool map, 27-tool confirmation list, phase-transition approval, per-skill OOB toggles, read-only Cypher with user_id+project_id filters, containerized execution.
Bounded Approval Workflow
Human-in-the-loop gates on Metasploit, Hydra, code execution and phase transitions. Flip a switch for fully autonomous mode.
Per-project Agent Behaviour, deterministic dangerous-tool list, stealth mode disables noisy tools and blocks live scanning.
Live Guidance Steering
Type a message while the agent is mid-engagement and it gets injected into the next reasoning step. Refocus, redirect or veto without stopping the session.
Examples: 'focus on SSH vulnerabilities', 'skip the web app, look at network services', 'try a different exploit module'. Acknowledged by the agent before the next think iteration.
Parallel Pentest Sessions
Run multiple independent agentic conversations per project, each with its own phase, todo list and execution trace. Switch between them without losing state.
Per-conversation checkpointing, days-later resume, agent-running indicators, auto-generated titles, Markdown export. Stop and Resume preserve full context.
Offline Security Knowledge Base
Local FAISS + Neo4j RAG index over curated security datasets (GTFOBins, LOLBAS, NVD, ExploitDB, OWASP, Nuclei, MITRE) queried before any external web search.
6-stage pipeline: hybrid retrieval, source boosts, cross-encoder rerank, MMR diversity, sufficiency check. ~250 ms vs 1-3 s Tavily, works offline, falls back automatically when confidence is low.
Natural Language → Cypher
Ask the agent “show endpoints with injectable params” and it generates the Cypher query, with 3-attempt error recovery.
25+ example patterns, automatic parameter binding, write ops blocked, tenant-filtered to prevent cross-project leakage.
Tradecraft Lookup
Curate trusted hacking sources (HackTricks, PayloadsAllTheThings). The agent fetches and cites pages mid-attack.
Per-user URL library, auto-fetch during chains, content embedded into reasoning steps with source attribution.
400+ LLM Models, 12 Providers
Swap the agent brain across OpenAI, Anthropic, OpenRouter, Google Gemini, DeepSeek, GLM, Kimi, Qwen, xAI Grok, Mistral, AWS Bedrock and any OpenAI-compatible endpoint, per project or per session.
Model-tuned system prompts, hot-swap, dynamic model fetching, OpenAI-compatible bucket covers Ollama, vLLM, LM Studio, Groq, Together AI, Fireworks AI, Deepinfra.
Built-in Agent Skills
Phase-aware workflows for CVE/MSF, SQLi, XSS, SSRF, RCE, Path Traversal/LFI/RFI, Hydra brute force, phishing simulation and availability testing. Auto-selected by the Intent Router from a single user message.
Per-skill tool routing (sqlmap, dalfox, commix, sstimap, ysoserial, hydra, msfvenom, slowhttptest), OWASP 4-stage rigor, OOB oracles via interactsh, RoE-gated activation.
User Agent Skills
Drop-in Markdown attack workflows from the community: XXE, BFLA, IDOR/BOLA, SSTI, insecure deserialization, mass assignment, subdomain takeover, insecure file uploads, advanced API/SQLi/XSS exploitation.
Each skill is a single .md file with phased steps and a description for classification, injected verbatim across all three phases. Upload your own from Global Settings, toggled per project.
Chat Skills
Reference library for tool flags, vulnerability theory and framework quirks. Inject any of them mid-conversation with /skill ffuf, /skill nextjs, /skill jwt, etc.
Categories cover Active Directory, Cloud, Post-Exploitation, Tooling, Protocols, API Security, Technologies, Frameworks, Vulnerabilities. Persist across messages, swap on demand, no impact on classification.
Agent Workspace
An interactive working environment with three editors sharing the same files: you through the live drawer, the agent through 24 filesystem tools, and the Kali terminal inside the sandbox. The agent uses it to draft payload lists, take scratchpad notes between phases, run scans in the background, and consolidate findings into reports that survive every turn of the engagement. The agent writes, you watch, the next prompt picks up where the last one left off.
read · write · edit (with undo) · grep · glob · find · tree · symbols · hash · diff · archive · extract
spawn long scans as detached async tasks · run 4+ in parallel · fs_grep their logs mid-flight
outputs over 20 KB land in tool-outputs/ with a head/tail stub · context never bloats
UI drawer + agent tool calls + Kali terminal · same files, instantly synced
One Agent, Twelve Brains
Hot-swap the model that drives the agent — per project, per session, mid-engagement. Frontier APIs, Chinese open-weights, your own local Ollama box. Tilt a card, click through to grab a key.
OpenAI
Frontier reasoning
Anthropic
Tool-use champion
OpenRouter
200+ aggregated
Google Gemini
Long context
DeepSeek
Cheap + open
GLM (Zhipu)
Multilingual
Kimi (Moonshot)
200k context
Qwen (Alibaba)
Open-source reasoning
xAI Grok
Real-time web
Mistral AI
European open-weight
AWS Bedrock
Enterprise foundation
OpenAI-Compatible
Ollama · vLLM · Groq
400+ models · dynamic fetching · OpenAI-compatible bucket covers Ollama, vLLM, LM Studio, Groq, Together AI, Fireworks AI, Deepinfra
From Terminal Chaos to Knowledge Graph
Drag the slider. On the left: the way pentest data lives without RedAmon — JSONL files, scrollback buffers, scattered notes. On the right: the same engagement, normalized into a Neo4j attack-surface graph the agent reasons over.
Drag the divider · same data · two ways of living with it
Recon as a Living Knowledge Graph
Industry-standard scanners chained so each tool's output feeds the next, then merged into a single Neo4j knowledge graph. Findings are deduplicated, relationships are explicit, and the agent inherits a structured, fully connected attack surface ready to query in natural language.
Recon Pipeline Modules
Each module is an independent stage of the deterministic recon engine. Findings flow downstream as graph nodes and relationships, ready for the agent to query.
Tool Output → Agent Knowledge
The modules are not isolated scanners. Each tool's output is the next tool's input: subdomains feed port scans, ports feed HTTP probes, URLs feed crawlers, JS bundles feed secret detectors. Every finding lands in the graph as a node, every dependency as a relationship.
Producer / consumer mapping locks every stage to the next, MERGE-based writes deduplicate across runs, and the resulting graph becomes the structured knowledge the agent reasons over via query_graph and Cypher.
17 Parallel Recon Modules
Independent modules running concurrently end-to-end: subdomain discovery, OSINT enrichment, port scanning, HTTP probing, web crawling, JS analysis, GraphQL testing, vhost discovery, takeover detection, vulnerability scanning, secret hunting and security configuration checks.
ThreadPoolExecutor fan-out, MERGE-based graph writes, RoE-aware tool gating, partial recon to re-run any single tool without redoing the pipeline.
80+ Tools, Libraries and Sources
Battle-tested CLI tools, APIs and detector libraries shipped across the modules: Subfinder, Amass, Nuclei, Httpx, Naabu, Masscan, Nmap, Katana, FFuf, jsluice, GVM, TruffleHog, Subjack, BadDNS, vulnx and 60+ more.
Full ProjectDiscovery suite, OWASP Amass, OpenVAS NVT corpus, vulnx + NVD + KEV + EPSS + MITRE ATT&CK / CWE / CAPEC, 7 passive OSINT sources, 90+ JS secret patterns, 700+ TruffleHog detectors, 40+ takeover provider fingerprints.
32-Node Attack Surface Graph
Every recon stage feeds a single Neo4j knowledge graph the agent reasons over: Domain → Subdomain → IP → Port → Service → CVE → Endpoint → Vuln.
32 node types, 50+ relationships, multi-tenant scoped (user_id + project_id on every node), EvoGraph bridges to attack chains.
JS Deep Reconnaissance
Extract API endpoints, cloud keys, JWT secrets, source maps and dependency confusion vectors from JavaScript bundles.
jsluice + 90+ regex patterns (AWS, GCP, Stripe, Twilio), framework fingerprinting, graph-indexed for agent reuse.
Every Tool Fully Configurable
Each module exposes its full parameter surface in the project settings UI: thread counts, rate limits, scan profiles, wordlists, severity filters, custom flags. UI fields become the tool's CLI arguments at scan time.
169+ project settings across dedicated per-tool forms, sensible defaults, validated input, immutable post-apply for reproducible scans, free-form custom-args field for power users.
170k+ NVT Network Vuln Scanning
Real network-level GVM/OpenVAS scans against live IPs (not just DAST) with QoD scoring and CVSS-rated findings.
Parallel container, configurable scan profiles (Full+fast, Full+slow), CVE/Vulnerability node feeds linked to IP/Port/Service.
TruffleHog 700+ Detectors
Scan repos for 700+ secret patterns and live-verify they actually authenticate against the issuing API.
Regex + entropy detection, optional API verification, verified flag on Secret nodes, supports private repos.
GitHub Secret Hunter
Mine GitHub search, repos and commit history for hardcoded credentials with 40+ entropy-validated patterns.
Public + private repos, full commit history, PDCP-compatible, results merged into the graph as Secret nodes.
VHost & SNI Hidden Host Discovery
Find admin panels, staging sites and k8s ingress backends hidden behind reverse proxies via L7 Host + L4 SNI probing.
Per IP×subdomain pair, separate BaseURL nodes, detection method tagged, exposes Cloudflare routing inconsistencies.
21 Recon Presets + AI-Generated
One-click profiles for API, web, internal, cloud or IoT. Or ask the AI to generate a custom preset for your target.
328+ pipeline parameters, immutable post-apply for reproducible scans, AI generation via LLM on demand.
Workflow View + Chain-Break Detection
Visual 11-stage pipeline diagram with animated edges and amber/red warnings when tool dependencies are starved.
Three-band layout (data / tools / data), click-through tool config modal, Tab/Workflow toggle, real-time edge animation.
Partial Recon Single-Tool Runner
Re-run any tool (Nuclei, Naabu, Katana) on custom inputs without redoing the full pipeline. Results MERGE into the graph with no duplicates.
Per-tool play button, custom-targets modal, MERGE logic with orphan UserInput cleanup, saved tool settings honoured.
AI in Pipeline
AI layered over the deterministic recon boosts and tunes every tool execution against the live response fingerprint, dropping wasted probes and sharpening coverage on modern targets.
Optional per-tool hooks, master toggle in Target tab, fail-safe fallback to static config on any LLM error.
CypherFix Closes the Vulnerability Loop
Two specialized AI agents pick up where reconnaissance and exploitation leave off. The triage agent reads the graph, dedupes findings across scanners, prioritizes by exploitability and proposes ranked remediation. The CodeFix agent then implements those fixes against your real repository and ships them as a pull request.
From Recon to Pull Request
Nine steps from a blank scope to a merged pull request. The recon pipeline collects findings deterministically. The agent attacks autonomously. The platform auto-writes the report. CypherFix patches the code. All in the same loop.
Recon
Full Recon Pipeline
Seventeen modules run in parallel and chain producer-to-consumer: subdomain discovery, OSINT enrichment, port scanning, HTTP probing, web crawling, JS deep recon, GraphQL testing, vhost discovery, takeover detection, vulnerability scanning, secret hunting and security configuration checks.
Knowledge Graph
Every finding lands as a typed Neo4j node and every dependency as an explicit edge. Findings are deduplicated across scanners on the way in. The graph becomes the structured memory the AI agent reasons over with Cypher and natural language.
Engagement
Set the Goal
Tell the agent in plain English what you want pwned. The Intent Router classifies the request, picks the right attack skill, and the agent plans its own todo list. Rules of Engagement and approval gates keep everything inside scope.
Exploitation
A fireteam of specialist sub-agents attacks in parallel: CVE / Metasploit, SQLi, XSS, SSRF, RCE, Path Traversal, Hydra brute force, phishing simulation. Each runs its own ReAct loop with full graph awareness and tool autonomy.
Post-Exploitation
Stateful Meterpreter sessions, lateral movement, privilege escalation, persistence and exfiltration. NAT-transparent reverse shells over ngrok or Chisel auto-tunnel from the Kali sandbox so the operator never has to expose ports.
Closing the Loop
Pentest Report
A six-section HTML report writes itself from the graph: executive summary, scope, risk overview, findings with CVSS / CVE / CWE / CAPEC attribution, attack surface map and remediation recommendations. Self-contained, client-ready, prints cleanly to PDF.
AI Triage
Hundreds of findings get deduplicated across scanners, prioritized by EPSS, KEV and CVSS, and turned into a ranked remediation plan. Cross-specialist convergence attribution shows which fireteam member found what.
CodeFix Agent
A second autonomous agent clones your repository, navigates the codebase with eleven code-aware tools (read, grep, edit, bash, github_*), implements the targeted fix and writes a clean diff with rationale.
GitHub Pull Request
A complete pull request lands on your repo with the fix, the diff, the rationale and the linked vulnerability ID. Review, request changes, or merge — the loop closes the moment the PR is approved.
Everything Else in the Box
Beyond the agent and the recon engine: governance, surface slicing, dashboards, training labs, knowledge base, reports, exports and multi-tenancy. The full toolkit a security team needs around an autonomous pentester.
Surface Shaper
Describe an attack surface in plain English (“CVEs CVSS>9 affecting our stack”). AI generates a read-only Cypher slice and saves named surfaces.
20 example queries, scopes both agent + canvas, infrastructure / vulnerability / web-app / threat-intel categories, reusable across sessions.
RoE Parsing & 3-Layer Enforcement
Upload a Rules-of-Engagement PDF. AI parses it into structured constraints, enforced at agent prompt + tool executor + pipeline filter.
Three-layer enforcement, deterministic hard gate, scope / exclusions / time windows / rate limits / compliance frameworks; non-bypassable by prompt injection.
Red Zone 2D / 3D Graph Canvas
Explore your full attack surface on a force-directed 2D or WebGL 3D graph, with real-time updates during scans and a click-through node drawer.
Node coloring per type, filter + legend, live properties drawer, pan/zoom/rotate, agent + scan toolbar controls.
16 Data Table Views
Switch from graph to table mode and slice the findings by type: Node Inspector, All Nodes, JS Recon, plus 13 dedicated finding tables (Vulnerabilities, CVEs, Subdomains, Endpoints, Secrets, Certificates, Ports, Services, Headers and more). Search, sort and export every one.
16 presets (3 base + 13 finding tables), per-table search and refresh, XLSX / JSON / Markdown export buttons on every preset, Node Inspector for property-level node drill-down.
HackLab: 48 Autonomous Scenarios
Proof your agent can actually pwn things: 48 scenarios across SQLi, NoSQL, RCE, XXE, SSRF, JWT, IDOR, prototype pollution.
Decoded walkthroughs, video demos, community submissions welcome, 12 categories on a deliberately vulnerable DVWS-Node + CVE lab.
Insights Dashboard Risk Analytics
Single-pane KPIs: 6-axis security radar, composite risk score 0 to 100, trend charts across the whole instance.
12 weighted signals (CVSS, EPSS, KEV, CAPEC, GVM, secrets), log-normalized radar prevents saturation, real-time aggregation.
CVE Intel Exploitability Engine
Query CVEs by CVSS≥8, EPSS≥0.5, KEV status and Nuclei template availability via Lucene syntax, with no manual KEV lookups.
vulnx CLI wrapper across NVD + KEV + EPSS + HackerOne + GitHub PoCs + Nuclei + CPE; 69 searchable fields, 6h refresh.
Offline Knowledge Base + Web Fallback
Agent queries a local KB (GTFOBins, LOLBAS, NVD, ExploitDB, OWASP, Nuclei) first. Web search runs only when KB confidence is low.
FAISS vector index + Neo4j fulltext, ≥0.35 confidence skips Tavily, works zero-API-key for sensitive engagements.
Browser-Based Kali Terminal
Full interactive xterm.js shell into the Kali sandbox container, with all 70+ pentesting tools pre-installed and no SSH needed.
WebSocket PTY (port 8016), signal handling, color support, agent proxy → kali-sandbox terminal server → /bin/bash.
Pentest Report Templates
Export findings as templated PDF / DOCX / HTML reports with executive summary, severity filters and embedded graph visualizations.
Multiple templates (executive, technical, compliance), auto-TOC, tool-output appendices, raw Cypher logs.
Data Export & Import
16 export presets covering Node Inspector, all nodes, JS Recon and 13 finding tables, in XLSX, JSON or Markdown.
Styled headers, per-column filters preserved, hyperlinks, full-project or per-table; programmatic JSON for downstream tools.
Multi-Tenant Project Management
Multi-user / multi-project tenancy with admin & standard roles, 169+ project settings and global API key rotation.
DB-level user_id + project_id filtering, per-project LLM provider, key rotation for rate-limit management, shared key vault.
See It In Action
Real RedAmon engagements recorded end to end. The agent picks targets, runs tools, adapts to findings and pwns the target while you watch. Unscripted, full timeline, no edits.