wgpsec/AboutSecurity is a large structured penetration-testing knowledge base that packages methodologies, dictionaries, payloads, and vulnerability data in formats an authorized AI agent like Claude Code can discover and invoke.
| Tool | wgpsec/AboutSecurity — AI Agent-executable pentest knowledge base (Skills, Dic, Payload, Vuln layers) |
| Category | Structured security knowledge base / agent skill repository |
| Primary Use | Syncing SKILL.md methodologies into .claude/skills/ so AI agents load tested workflows during authorized assessments, and querying dictionaries/payloads via the context1337 MCP service |
| Safe Use | Educational reference and methodology scaffolding for penetration testers working under explicit written authorization; equally valuable to defenders mapping attacker tradecraft and building detection content |
| Telemetry Note | The repository is pure data — no runtime telemetry of its own; the attack techniques it documents (e.g., postexploit/ persistence patterns) are what defenders should baseline and detect in authorized lab environments |
wgpsec/AboutSecurity answers a question that has become urgent in the agentic-security era: how do you hand an AI coding or testing agent two decades of pentest tradecraft without it hallucinating methodology? The project's answer is structure. Rather than a sprawling wiki, it organizes everything into four machine-consumable layers — Skills/, Dic/, Payload/, and Vuln/ — each with naming conventions and _meta.yaml metadata designed specifically so that tools like Claude Code can discover, index, and invoke the right resource from conversational context. At roughly 1.7k stars and written primarily in Python, it is positioned as the knowledge layer of the broader WgpSec Agentic Ecosystem.
The Skills/ directory is the heart of the project: over 200 methodology files, each a SKILL.md describing "what to do when encountering scenario X." The taxonomy is broad and deliberately full-chain. recon/ covers subdomain enumeration and passive information gathering; exploit/ organizes web methodology (web-method/), authentication attacks (auth/), network-service testing by port and protocol, binary exploitation, and advanced categories like HTTP request smuggling and race conditions; lateral/ documents active-directory attack patterns including NTLM relay and Kerberoasting. For auditors, code-audit/ ships dedicated PHP and Java 8-skill systems covering injection, file handling, serialization, auth, and framework-specific exploit chains.
Several categories are notable for their currency. ai-security/ covers prompt injection, model jailbreaking, prompt leaking, and agent attack chains — a body of knowledge most traditional cheat sheets simply lack. cloud/ addresses Docker escape, K8s attack chains, AWS IAM misconfigurations, Alibaba and Tencent Cloud specifics, and serverless. dfir/ and threat-intel/ round out the defensive side with memory forensics, anti-forensics, and threat-hunting material, which makes the repo genuinely dual-use rather than purely offensive.
The Vuln/ layer contains 600+ structured vulnerability entries organized by product: ai/ (ComfyUI, Dify, LangFlow, AnythingLLM), cloud/, middleware/ (437 entries spanning ActiveMQ, Nacos, Grafana, Jenkins, RocketMQ), network/, and web/ applications. The README draws an important architectural distinction: Skills/ documents the post-exploitation layer — what to do after access is gained — while Vuln/ is the vulnerability data layer holding affected versions and per-CVE detail. This separation of "how you got in" from "what you do once inside" is clean conceptual hygiene, and it mirrors how mature assessment teams structure their own internal runbooks.
The Dic/ and Payload/ layers are brute-force wordlists and categorized payload collections, but the engineering detail is what distinguishes them from a recycled SecLists mirror. Every directory carries a _meta.yaml with category, description, per-file line counts, usage notes, and bilingual tags — effectively a built-in index for AI search. Dictionaries cover 19 service types (mysql, redis, ssh), device default credentials, web directories, API parameters, and upload-bypass lists; payloads span sqli/, xss/, ssrf/, xxe/, rce/, cors/, and prompt-injection/, all in lowercase-hyphenated directories. For defenders, this metadata is a gift: it is effectively a labeled taxonomy of attacker tooling inputs.
Consumption happens in two ways. The first is the sync script: ./scripts/sync-claude-skills.sh --target /path/to/your-project creates .claude/skills/<skill-name>/ symlinks mapping the repo's nested category tree into the flat structure Claude Code recognizes, after which the agent automatically matches relevant skills to conversation context. The second, which the authors recommend, is context1337 — a standalone MCP server that turns the repo into a searchable natural-language API (claude mcp add aboutsecurity --transport http http://localhost:1337/mcp). Dictionaries and payloads need no syncing at all; you simply reference their paths in conversation and the agent reads them via its file tools.
The ecosystem framing is ambitious and worth understanding before you invest. AboutSecurity feeds context1337 (the service layer), which in turn feeds tchkiller, an autonomous pentest agent with multi-round decision-making, evaluated against benchmark-platform, a CTF range distributed via GitHub Releases. A private general-purpose engine called PoJun sits underneath. The README is transparent about lineage, citing Anthropic's own skill-creator spec plus community projects like hacktricks, InternalAllTheThings, and several audit-skill repos — a healthy sign that this is consolidation and structuring of community knowledge rather than wholesale invention.
From a governance standpoint, the repo's contribution standards matter. CONTRIBUTING.md specifies the skill format, Vuln database writing standards, reference requirements, and a benchmark testing process — meaning skills are expected to be evaluated against the CTF range before acceptance. That is a materially higher bar than the average cheat-sheet repository, and it addresses the core trust problem with agent-consumed security knowledge: an agent acting on untested methodology is worse than no methodology at all.
For authorized professionals, the practical value is triage speed. A consultant scoping an engagement against a middleware-heavy stack can pull the relevant Vuln/middleware/ entries in seconds through context1337; a code-audit team gets a standardized PHP/Java skill scaffold instead of tribal knowledge; purple teams can mine postexploit/ and evasion/ categories to build detection content for persistence and C2 patterns before adversaries exercise them. Defenders get equal mileage: the structured payload and dictionary taxonomies map directly onto WAF rule tuning and credential-policy hardening work.
Watch a few things. The _meta.yaml descriptions are documented as Chinese with bilingual tags, so English-only teams should expect mixed-language content in places. There is no license declared in the repository metadata at the time of writing, which matters if you plan to redistribute or fork the corpus commercially. And because the repo catalogs genuinely dangerous categories — tool-delivery/, persist-maintain/, evasion — it belongs in the hands of credentialed professionals operating in labs and contracted engagements, not as an autonomous agent's unsupervised bedtime reading. Used that way, it is one of the more thoughtfully engineered knowledge bases in the agentic-security space.
wgpsec/AboutSecurity.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.