SecSkills is a link directory of AI security agents, not a tool you can run

The repository is a README, a licence and an image, and that is the product

SecSkills collects and organises Agents, Skills and MCP servers related to penetration testing, vulnerability scanning, code audit, CTF, reverse engineering and security research. The description is in Chinese, and the whole catalogue is written that way.

The repository contents are worth stating plainly, because it determines what you can do with it. The top level holds LICENSE, README.md and gzh.png. There is no source directory, no package manifest, no build configuration, no test suite and no script of any kind. Everything the project offers is the README itself, which is a set of Markdown tables of links.

That makes it a directory rather than a tool. There is nothing to install, nothing to configure and nothing to run. The value is in the organisation: someone has already gone through the ecosystem, written a one-line description of each entry, and sorted the results into categories with counts. Whether that curation is worth your time depends on whether you were going to do the same search yourself, and the descriptions do include enough detail to tell a promising project from a duplicate.

Seventeen categories, and the category counts show where the field is crowded

The catalogue header states a total of 102 entries and lists 17 categories with their counts, which is a map of the field rather than a preference list. Agent is the largest at 46. Penetration testing and vulnerability scanning holds 16, code audit holds 12, and entertainment holds 12, which covers the CTF and challenge-solving side. Then the single-entry categories: sample analysis, incident response, CTF, red and blue teaming, reverse engineering, report writing, domain penetration, WeChat mini program audit, MCP, and security research. JS reverse engineering and skills checking have 2 each, mobile security has 2.

Two of those numbers deserve a second look. The MCP category has exactly one entry, so if you came looking for Model Context Protocol servers as a category, the interesting material is actually spread across the Agent table, where several entries are described as MCP servers. The second is that skills checking has two, which is a self-check angle rather than a capability angle: the catalogue includes a way to inspect the skills you already have, not just ways to obtain new ones.

The code audit section is where the entries get most specific. There is a Java code audit skill collection described as focused on Java code audit with automated source analysis, route extraction and parameter mapping. There is a PHP Code Audit Skill covering a full flow from route enumeration through authorisation modelling, data flow tracing and classified vulnerability auditing to evidence consistency checking and report summarisation, implemented in environments such as Cursor by executing the Agent according to the SKILL document. And there is a multi-agent audit framework built on Claude Code Agent Teams, covering environment construction, static reconnaissance, dynamic tracing, deep adversarial exploitation, post-exploitation correlation analysis and report closing, supporting 21 vulnerability types.

The Agent table mixes live projects with archived ones, and says which is which

The Agent table lists 46 entries, and the descriptions carry the kind of detail you would want before clicking. strix is described as an open source AI hacker that finds and fixes application vulnerabilities, shannon as an autonomous white box AI penetration testing tool for web applications and APIs, and pentagi as a fully autonomous AI agent system for complex penetration testing tasks. PentestGPT is labelled the early benchmark in the LLM-driven automated penetration testing agent space.

Some entries are MCP servers rather than agents, and this is where the categories overlap. hexstrike-ai is an MCP server letting AI agents run 150+ security tools on their own. pentest-ai is an MCP server wrapping 205+ security tools, 17 specialist agents and deterministic vulnerability verification with a zero false positive claim, offering both a CLI and an MCP path with its own LLM. Dark-Moon covers web, cloud, AD and K8s with multi-agent orchestration, a privacy gateway and 50+ tool integrations. T3MP3ST is an autonomous red team and multi-agent offensive security meta framework that reuses local AI coding agents such as Claude Code, Codex and Ollama as a zero-day hunter.

The catalogue is honest about the lifecycle. cai, the Cybersecurity AI framework, is marked as archived on 2026-08-28, having produced 18 papers and 30+ CVEs. Cyber-Zero, a training-free runtime cybersecurity agent from Amazon Science, is marked as archived on 2026-07-10. Cyber-AutoAgent is marked as archived but representative. BugTrace-AI is marked as archived and having evolved into BugTraceAI v2. That is useful curation, because archived research code is often still worth reading while being useless to depend on.

Several entries cite measurable results, and those are the claims of the listed projects rather than of the catalogue: ctf-agent as first place at BSidesSF 2026, BreachWeave as first place out of 613 in the second Tencent Cloud hackathon final, ctf-agent and a couple of others with XBOW verification scores, BoxPwnr as a benchmark framework across 15 platforms, HackSynth with a planner and summarizer module over 200 PicoCTF and OverTheWire challenges, and deadend-cli with an XBOW black box score of 81% at about 122 dollars of API cost running locally.

Code audit and reverse engineering entries describe workflow, not tools

The more specialised categories read differently from the Agent table. The code audit entries are described as process pipelines rather than as point tools, and the PHP entries are the clearest example: one lays out route enumeration, authorisation modelling, data flow tracing, classified vulnerability auditing, evidence consistency checking and report summarisation as a sequence, and the other frames the same kind of chain across environment construction, static reconnaissance, dynamic tracing, deep adversarial exploitation, post-exploitation correlation analysis and report closing.

That distinction matters if you are evaluating this for a team. A pipeline skill set that insists on an evidence consistency check before the report is a different kind of artifact from a scanner that prints findings. It imposes a review sequence, and it presumes you have a way to execute Agent-style steps in your environment.

The JS reverse engineering and mobile security categories are small, two entries each, and the sample analysis, incident response, domain penetration, WeChat mini program audit, report writing and red and blue teaming categories have one entry apiece. The WeChat mini program audit slot is worth naming because it is a genuinely specific niche that a general security agent catalogue usually skips.

There is also a separate MCP category, and the entertainment category with 12 entries is separate from the single CTF entry, which suggests the catalogue treats challenge tooling and entertainment material as different audiences.

The licence and the promotion block are the two things to look at first

SecSkills is licensed under Apache-2.0, which covers the README text, the table structure and the image. It does not extend to any of the linked projects, each of which carries its own terms, and the catalogue itself is a set of links to GitHub repositories. Anyone adopting one of the listed agents inherits that project's licence and its security posture, not Apache-2.0.

The second thing at the top of the README is worth flagging, and it is easy to miss because it sits above the project title. The document opens with a section recommending a model relay service, ai.daoyivip.com, noting that registration yields a balance and that models are discounted from a tenth of the usual price. That is an advertisement inside the document, placed before the catalogue content.

It does not affect the accuracy of the tables, but it is worth knowing that the document is also a promotional surface, which is a reason to treat the entries as pointers rather than as endorsements. The descriptions contain performance claims for the listed projects, and those claims belong to the projects, not to the catalogue.

The last push to the repository was on 2026-09-14, and there are no GitHub releases, so the catalogue has no versioned history. A snapshot of the ecosystem in one file, with no way to see how a given entry changed over time, which is a real limitation for anything you plan to keep referencing.

What a directory cannot do for you

The limitation is structural rather than fixable by patching. A catalogue tells you that hexstrike-ai exists and that it exposes 150+ security tools over MCP, but it tells you nothing about how to install it, which node versions it needs, whether its output is trustworthy on your target, or whether it is compatible with the agent runtime you already use. None of that is in the README, and none of it can be, because the catalogue is one file and each linked project is a separate repository with its own documentation.

That also means the entries can rot. The catalogue already marks four projects as archived, which is evidence that the links do age, and it cannot tell you about the ones that were never popular enough to be listed but are now the better option. There is no CI, no link checker, no automated refresh and no version pinning anywhere in the repository, so the accuracy of any given entry is a function of when a human last looked at it.

There is also a scope question worth settling before you spend time here. The catalogue is explicitly about AI agents, skills and MCP servers for security work, and the tables are dominated by autonomous penetration testing agents. If your need is a static analysis tool, a network scanner or a reverse engineering suite that runs on its own, the mainstream tools in those categories already exist and are better documented than anything an agent framework wraps. The catalogue is at its most useful for the parts of the field where the tooling is genuinely new and hard to find, which is exactly the AI agent and MCP layer.

Editorial conclusion

SecSkills fits a security engineer or researcher who wants an overview of what exists in the AI security tooling space and wants the original repository links rather than a curated toolchain, and it does not fit a team looking for an installable, versioned security framework or a dependency you can pin. Before relying on it, follow each entry to its own repository and read that project's documentation, because the catalogue itself carries no install steps, no configuration and no compatibility information, and several of the listed projects are already archived.

Frequently asked questions

Official sources

1. DaoYiSec/SecSkills on GitHub
2. Issues
3. License: Apache-2.0
4. README