Multi-platform AI Agent security assessment tool — supports 9 major platforms including OpenClaw / Dify / FastGPT / LobeChat
78 DAG attack chains | 78 Exploit modules | 9 platforms supported | ATT&CK stage mapping | Async Tokio engine | Attack graph visualization
English · Japanese · Русский · Deutsch · Français
⚠️ COMMERCIAL USE STRICTLY PROHIBITEDThis project is licensed under the CatchClaw Strict Non-Commercial License v2.0.
Any form of commercial use is strictly prohibited without written authorization from the copyright holder (Coff0xc). Violators will be held legally liable.
Prohibited activities include, but are not limited to:
- Selling, sublicensing, or renting this software or its derivative works
- Using this software for SaaS, penetration testing services, consulting, or any paid service
- Integrating it into commercial products, platforms, or tools
- Using it to train commercial AI/ML models
- Renaming, reskinning, or repackaging and redistributing it
- Any activity that directly or indirectly generates revenue
The copyright holder reserves the right to pursue claims indefinitely, including recovery of all profits, legal fees, and punitive damages.
See LICENSE for details.
┌────────────────────────────────────────────────────────────────────────────┐
│ CatchClaw v5.3.0 │
├────────────────────────────────────────────────────────────────────────────┤
│ ● 78 DAG attack chains ● 78 Exploit modules ● Async Tokio engine │
│ ● ATT&CK 9-stage mapping ● Mermaid graph export ● JSON/HTML/MD reports │
│ ● Kahn topological sort ● Semaphore control ● Conditional/fallback │
│ ● Multi-target scan (CIDR) ● Port scan/service discovery ● 200+ Payloads│
│ ● 9 platforms ● Multi-platform fingerprinting ● --platform auto-detect │
├────────────────────────────────────────────────────────────────────────────┤
│ Attack surface: WS API | HTTP REST | OAuth | Webhook | Node | Multi-Platform │
│ Coverage: SSRF | RCE | key theft | hijacking | privesc | persistence | leaks │
│ New: C2 exfil, Skill poison, Agent inject, MCP inject, OAuth theft, DNS Rebind │
└────────────────────────────────────────────────────────────────────────────┘
- Project Overview
- Core Features
- Supported Platforms
- Installation
- Quick Start
- CLI Usage
- 78 Exploit Modules
- DAG Attack Chain Architecture
- Nuclei Templates
- Project Structure
- Disclaimer
CatchClaw is a Rust-based multi-platform AI Agent security assessment tool supporting 9 major AI platforms, including OpenClaw / Open-WebUI, Dify, FastGPT, and LobeChat. With 78 DAG attack chains and 78 Exploit modules, it covers the full ATT&CK kill chain, from initial reconnaissance to data exfiltration.
Built on the Tokio async runtime, the DAG engine executes level by level in parallel using Kahn's topological sort, with a Semaphore controlling concurrency, and supports conditional execution and fallback nodes. Attack graphs can be exported as Mermaid flowcharts.
| Scenario | Manual Testing | CatchClaw |
|---|---|---|
| Vulnerability validation | Hand-written PoCs, one by one | 78 chains orchestrated and validated automatically |
| Attack surface coverage | Based on experience | Full coverage of WS + HTTP + OAuth + Webhook + Node |
| Dependency chain discovery | Hard to trace | DAG automatically discovers vulnerability dependency paths |
| Result visualization | Organized by hand | Mermaid attack graphs + JSON/HTML/Markdown reports |
| CI/CD integration | None | 24 Nuclei templates, plug and play |
| Multi-target scanning | Tested manually, one by one | Batch scanning of CIDR blocks, IP ranges, and target files |
| Service discovery | Requires nmap | Built-in port scanning + multi-platform fingerprinting |
| Multi-platform coverage | Each platform researched separately | Automatic detection across 9 platforms: OpenClaw/Dify/FastGPT/NextChat/AnythingLLM/Flowise/RagFlow/LibreChat/LobeChat |
|
|
|
|
| Platform | Version | Fingerprinting | CVE Coverage | Probe Module |
|---|---|---|---|---|
| OpenClaw / Open-WebUI | — | Auto | 72 dedicated modules | Built-in |
| Dify | v1.5+ | Auto | CVE-2025-56157, CVE-2025-55182 | dify_probe |
| FastGPT | v4.9+ | Auto | CVE-2025-27600, CVE-2025-52552 | fastgpt_probe |
| NextChat | v2.12+ | Auto | CVE-2023-49785 | nextchat_probe |
| AnythingLLM | v1.3+ | Auto | CVE-2024-13059, CVE-2024-0455 | anythingllm_probe |
| Flowise | v1.6+ | Auto | CVE-2025-26319, CVE-2025-58434 | flowise_probe |
| RagFlow | v0.14+ | Auto | CVE-2024-12433, CVE-2024-12450 | ragflow_probe |
| LibreChat | — | Auto | CVE-2025-69222, CVE-2025-69220 | librechat_probe |
| LobeChat | — | Auto | CVE-2026-23733 | lobechat_probe |
git clone https://github.com/Coff0xc/catchclaw.git
cd catchclaw/rust
# Debug build
cargo build
# Release build (optimized, with symbol table stripped)
cargo build --release
# The binary is at target/release/catchclaw- Rust Edition 2024 (rustc 1.85+)
- Windows / Linux / macOS supported
# List all registered modules
catchclaw list
# Auto-detect the platform type
catchclaw scan -t TARGET_IP:PORT
# Specify the platform type
catchclaw scan -t TARGET_IP:PORT --platform dify
# Scan and output a JSON report
catchclaw scan -t TARGET_IP:PORT -o report.json
# Output an HTML report
catchclaw scan -t TARGET_IP:PORT -o report.html --format html
# Multi-target scan (CIDR)
catchclaw scan --targets "192.168.1.0/24:8080"
# Batch scan from a file
catchclaw scan -f targets.txt -o results.json
# Use a Profile preset
catchclaw scan -t TARGET_IP:PORT --profile stealth
# Show only critical/high-severity results
catchclaw scan -t TARGET_IP:PORT --severity-filter critical,high
# Preview the execution plan
catchclaw scan -t TARGET_IP:PORT --dry-run
# Scan with a token
catchclaw scan -t TARGET_IP:PORT --token "your-gateway-token"
# Run the full attack chain
catchclaw exploit -t TARGET_IP:PORT --token xxx
# Run a single attack chain
catchclaw exploit -t TARGET_IP:PORT --chain-id 30CatchClaw v5.3.0 — Multi-platform AI Agent security assessment tool
Usage: catchclaw <COMMAND>
Commands:
scan Full security scan (build DAG → execute → summarize)
exploit Run attack chains (full DAG or a single node)
list List all registered Exploit modules
config Show the current configuration
Scan Flags:
-t, --target <HOST:PORT> Target address (single target)
--targets <SPEC> Multiple targets: comma-separated, CIDR, or IP range
-f, --targets-file <FILE> Target file (one host:port per line)
--token <TOKEN> Gateway Token (or CATCHCLAW_TOKEN environment variable)
--timeout <SECS> Request timeout in seconds (default 10)
-o, --output <FILE> Report output path
--format <FMT> Output format: json | html | markdown (default json)
--concurrency <N> Maximum concurrency (default 10)
--tls Use HTTPS/WSS
--callback <URL> SSRF callback URL
--profile <NAME> Use a Profile preset from the config file
--severity-filter <SEV> Filter by severity: critical,high,medium,low,info
--dry-run Preview the DAG execution plan
--config <FILE> Config file path (TOML/YAML/JSON)
--platform <PLATFORM> Specify the target platform type (auto|openclaw|dify|fastgpt|nextchat|anythingllm|flowise|ragflow|librechat|lobechat)
Exploit Flags:
-t, --target <HOST:PORT> Target address
--token <TOKEN> Gateway Token
--chain-id <ID> Run a single chain node ID
--concurrency <N> Maximum concurrency (default 10)
--tls Use HTTPS/WSS
Grouped by ATT&CK stage and attack category:
| Module | Category | Description |
|---|---|---|
| cors_bypass | Config | Origin reflection → cross-origin WS/API access |
| ws_hijack | Transport | Cross-origin WebSocket upgrade + token replay |
| auth_mode_abuse | Auth | Authentication mode detection and abuse |
| log_disclosure | DataLeak | Credential/sensitive data leakage via logs.query |
| hidden_content | DataLeak | Hidden content discovery |
| origin_wildcard | Config | Origin wildcard configuration detection |
| librechat_probe | Recon | LibreChat security probe (CVE-2025-69222/69220/54868) |
| lobechat_probe | Recon | LobeChat security probe (CVE-2026-23733 XSS) |
| dify_probe | Recon | Dify security probe (CVE-2025-56157 default credentials, SSRF) |
| fastgpt_probe | Recon | FastGPT security probe (CVE-2025-27600 SSRF, XSS) |
| nextchat_probe | Recon | NextChat security probe (CVE-2023-49785 SSRF) |
| anythingllm_probe | Recon | AnythingLLM security probe (CVE-2024-0455 SSRF) |
| flowise_probe | Recon | Flowise security probe (CVE-2025-26319 file upload) |
| ragflow_probe | Recon | RagFlow security probe (CVE-2024-12433 RCE) |
| Module | Category | Description |
|---|---|---|
| ssrf | SSRF | browser.request/navigate → cloud metadata (AWS/GCP/Azure) |
| eval_inject | Injection | eval/exec code execution |
| prompt_inject | Injection | System prompt extraction + instruction override |
| mcp_inject | Injection | MCP plugin injection |
| pairing_brute | Auth | 6-digit DM pairing code brute force |
| oauth_abuse | Auth | Slack OAuth redirect hijacking |
| responses_exploit | API | /v1/responses authentication bypass + tool injection |
| ws_fuzz | Fuzz | Malformed JSON-RPC + method injection |
| acp_bypass | Auth | ACP access control bypass |
| ssrf_rebind | SSRF | DNS Rebind SSRF |
| ssrf_proxy_bypass | SSRF | Proxy bypass SSRF |
| browser_request | SSRF | Browser request internal dispatch |
| csrf_no_origin | Config | CSRF without Origin validation |
| Module | Category | Description |
|---|---|---|
| apikey_steal | Credential | Provider API key extraction |
| oauth_token_theft | Credential | OAuth token theft |
| secret_extract | Credential | Plaintext theft via secrets.list + secrets.get |
| secrets_resolve | Credential | secrets.resolve internal injection API |
| talk_secrets | Credential | Key leakage via talk.config(includeSecrets) |
| Module | Category | Description |
|---|---|---|
| rce | RCE | Command execution probe via system.run |
| hook_inject | RCE | Command execution via webhook endpoint injection |
| tools_invoke | RCE | Bypass the Chat-layer security via tools.invoke |
| keychain_cmd_inject | RCE | Keychain command injection |
| qmd_cmd_inject | RCE | QMD command injection |
| exec_race_toctou | RCE | Execution race condition (TOCTOU) |
| exec_socket_leak | RCE | Execution socket leak |
| Module | Category | Description |
|---|---|---|
| agent_inject | Injection | Backdoor via agents.create/update + system prompt leakage |
| agent_file_inject | Injection | Persistent prompt backdoor via agents.files.set |
| channel_inject | Injection | Unsigned command injection for Mattermost/Slack/Discord |
| skill_poison | Injection | Skill poisoning |
| cron_bypass | Persistence | Cron blacklist bypass + persistence |
| session_file_write | RCE | Arbitrary file write via sessions.patch |
| patch_escape | RCE | Path traversal in apply_patch → arbitrary file write |
| link_template_inject | Injection | Link template injection |
| Module | Category | Description |
|---|---|---|
| approval_hijack | Auth | Prefix ID matching + execution policy tampering |
| config_tamper | Config | config.set security configuration write |
| rogue_node | RCE | Self-approved node pairing → command interception |
| silent_pair_abuse | Auth | Silent pairing abuse |
| auth_disable_leak | Auth | Authentication-disable leakage |
| Module | Category | Description |
|---|---|---|
| session_hijack | DataLeak | IDOR in sessions.preview + cross-session injection |
| transcript_theft | DataLeak | Session transcript theft |
| memory_data_leak | DataLeak | Memory data leakage |
| c2_exfil | DataLeak | C2 data exfiltration |
| browser_upload_traversal | RCE | Browser upload path traversal |
| secret_exec_abuse | RCE | Secret execution abuse |
| bypass_soul | Injection | Soul bypass |
| marker_spoof | Injection | Marker spoofing |
| redact_bypass | DataLeak | Redaction bypass |
| obfuscation_bypass | Injection | Obfuscation bypass |
| unicode_bypass | Injection | Unicode bypass |
| ratelimit_scope_bypass | Config | Rate-limit scope bypass |
| flood_guard_reset | Config | Flood guard reset |
| webhook_verify | Config | Webhook verification bypass |
| skill_scanner_bypass | Config | Skill scanner bypass |
| Module | CVE / Source | Description |
|---|---|---|
| gateway_hijack | CVE-2026-25253 | gatewayURL WebSocket hijacking → token theft |
| safebins_bypass | CVE-2026-28363 | GNU long-option abbreviation bypass of the safeBins allowlist → RCE |
| ws_auth_brute | CVE-2026-32025 | WebSocket authentication brute force + localhost rate-limit exemption |
| localhost_trust | ClawJacked | Authentication bypass via implicit localhost trust |
| guest_mode_abuse | Conscia audit | Guest Mode exposes dangerous API endpoints |
| mdns_leak | Conscia audit | mDNS/HTTP configuration parameter leakage |
| skill_supply_chain | ClawHavoc | Malicious Skill supply-chain attack detection (AMOS credential theft) |
┌─────────────────────┐
│ Level 0 (Recon) │
│ CORS / WS / Auth │
└─────────┬───────────┘
│
┌───────────────────┼───────────────────┐
▼ ▼ ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Level 1 (Init) │ │ Level 1 (Cred) │ │ Level 1 (Init) │
│ SSRF / Eval / │ │ APIKey / OAuth │ │ Prompt / MCP / │
│ Pairing │ │ Token Theft │ │ Responses │
└────────┬────────┘ └────────┬────────┘ └────────┬────────┘
│ │ │
└───────────────────┼───────────────────┘
▼
┌─────────────────────────┐
│ Level 2 (Execution) │
│ RCE / Hook / Tools / │
│ Session Write │
└────────────┬────────────┘
│
┌────────────┼────────────┐
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ Level 3 (Persis)│ │ Level 3 (Priv) │
│ Agent / Cron / │ │ Approval / Node │
│ Skill Poison │ │ Config Tamper │
└────────┬────────┘ └────────┬────────┘
│ │
└────────────┬────────────┘
▼
┌─────────────────────────┐
│ Level 4 (Exfil) │
│ C2 / Transcript / │
│ Memory Leak │
└─────────────────────────┘
Execution characteristics:
- Kahn topological sort → level-by-level parallel execution
Semaphorecontrols maximum concurrencydepends_on— predecessor nodes must completefallback_for— skip the fallback node if the parent node has findingscondition— conditional execution based on earlier resultsAttackGraph— records each node's executed/skipped/fallback status and exports a Mermaid diagram
nuclei-templates/ contains 24 standalone Nuclei-compatible YAML templates:
# Scan a single target
nuclei -t nuclei-templates/ -u http://TARGET_IP:PORT
# Scan a target list
nuclei -t nuclei-templates/ -l targets.txt
# Critical only
nuclei -t nuclei-templates/ -u http://TARGET_IP:PORT -severity criticalcatchclaw/
├── rust/
│ ├── Cargo.toml # Project configuration
│ └── src/
│ ├── main.rs # CLI entry point (clap derive)
│ ├── config/mod.rs # AppConfig + Profile + protocol constants
│ ├── chain/
│ │ ├── dag.rs # DAG engine (topological sort + concurrency + AttackGraph)
│ │ └── chains.rs # Definitions of the 78 attack chain nodes
│ ├── exploit/
│ │ ├── registry.rs # ExploitMeta + inventory registration system
│ │ ├── base.rs # ExploitCtx shared context + PayloadRegistry integration
│ │ └── *.rs # Implementations of the 78 exploit modules
│ ├── platform/
│ │ ├── mod.rs # PlatformProfile trait + TargetPlatform
│ │ ├── registry.rs # PlatformRegistry (9 platform registrations)
│ │ └── profiles/ # Implementations of the 9 platform profiles
│ ├── scan/mod.rs # Full scan + multi-target scan orchestration
│ ├── report/mod.rs # JSON / HTML / Markdown report output
│ └── utils/
│ ├── types.rs # Target / Finding / Severity / multi-target parsing
│ ├── http.rs # HTTP client + false-positive filter
│ ├── ws.rs # GatewayWsClient (WS + challenge)
│ ├── discovery.rs # Port scanning + multi-platform service discovery
│ ├── payload.rs # PayloadRegistry (YAML loading + directory merging)
│ └── mutate.rs # Payload mutation engine
├── payloads/ # 200+ external Payloads (YAML)
│ ├── ssrf.yaml # SSRF: AWS/GCP/Azure/IP bypass/protocol smuggling
│ ├── injection.yaml # Command injection/shell metacharacters/encoding bypass
│ ├── prompt_inject.yaml # Prompt injection/jailbreak/role override
│ ├── auth_bypass.yaml # Token/Header/path traversal/default credentials
│ └── xss.yaml # XSS: reflected/event-based/filter bypass/polyglot
├── nuclei-templates/ # 24 Nuclei YAML templates
├── LICENSE # CatchClaw Strict Non-Commercial License v2.0
└── README.md
This tool is intended for authorized security testing only. Test only systems you own or have explicit written authorization to test. Unauthorized access to computer systems is illegal. The author is not responsible for any misuse.
Coff0xc — https://github.com/Coff0xc
CatchClaw Strict Non-Commercial License v2.0 — commercial use is strictly prohibited; violators will be held liable.