“You know those rabbit holes you fall into when you’re supposed to be doing something entirely different? This was one of those.”
It started with my portfolio. Nothing dramatic — just me, late at night, rebuilding the thing from scratch using Claude Code, trying to add a security audit section that didn’t look like I’d copy-pasted it from a Stack Overflow answer from 2019.
I was browsing the Claude Code skill marketplace looking for something useful. Typed “security” into the search bar, half-expecting to find three mediocre options and move on.
What came back genuinely stopped me.
A repo called Anthropic-Cybersecurity-Skills, maintained by Mukul (GitHub: mukul975). 817 structured cybersecurity skills for AI coding agents. Mapped to six major industry frameworks. Covering 29 security domains. Featured at Black Hat USA 2026's Arsenal Lab — the curated, juried section of Black Hat reserved for genuinely useful open-source security tooling. Over 27,000 GitHub stars. Free. Apache 2.0.
The portfolio sat unfinished for the next hour while I read through everything this thing actually does.
What Even Is a “Cybersecurity Skill” for an AI Agent? 🧩
“Before I got excited, I had to understand what I was actually looking at — and it’s more interesting than it sounds.”
If you’ve read my 8-layer Claude Code setup piece, you’ll know I think carefully about giving Claude Code the right context for the right task. Skills are central to that — a skill is a SKILL.md file that teaches your AI agent exactly how to do something specific, invoked either by name (you type /security-review) or automatically when Claude detects the context is right.
Think of a skill less like a prompt and more like a playbook. It doesn’t just describe a concept — it tells the agent what technique it’s covering, what framework category it belongs to, what the context is (offense vs defense, detection vs mitigation), and how to actually execute it with enough specificity that the agent can do real work instead of talking around the topic.
Mukul built 817 of those. Each one follows the agentskills.io standard — the same format that works across Claude Code, GitHub Copilot, Cursor, Windsurf, Codex CLI, Gemini CLI, and 20+ other agent platforms. Install once, works everywhere you actually work.
Why Six Frameworks and Not Just One? 🗺️
“Anyone can make a list of security things. Mapping them to actual frameworks is where the real engineering work is.”
This is the part that stopped me from just skimming and made me genuinely read. Every single skill in this library simultaneously maps to six canonical security frameworks — and understanding why that’s significant requires knowing what each framework is actually for.
MITRE ATT&CK is the gold standard taxonomy for how attackers operate. It organizes techniques by kill chain phase: Initial Access, Execution, Persistence, Privilege Escalation, Defense Evasion, Credential Access, Discovery, Lateral Movement, Collection, Command and Control, Exfiltration, Impact. If you want to understand not just “SQL injection exists” but “SQL injection lives in the Initial Access and Execution phases, here’s how attackers chain it to later-stage access” — ATT&CK is the framework that makes that thinking structured and repeatable.
NIST CSF 2.0 is different. The National Institute of Standards and Technology Cybersecurity Framework organizes security around five functions: Identify, Protect, Detect, Respond, Recover. Where ATT&CK tells you what attackers do, NIST CSF tells you where in your organization’s lifecycle a given control or technique belongs. A CISO planning a security program and a developer picking security tools are asking different questions — NIST CSF answers the first.
MITRE ATLAS is the one that actually made me sit up. ATLAS maps adversarial tactics specific to AI and machine learning systems — not traditional software, but the specific attack surface of models, agents, and AI pipelines. The 2026 release covers 101 techniques and 77 sub-techniques, and there’s a whole category added in late 2025 specifically for agentic AI attack vectors: context poisoning, tool invocation abuse, MCP server compromises, and malicious agent deployment.
That last category is worth sitting with. I’ve written about Claude Code extensively. The workflows I describe — agents with access to my file system, my device, my codebase — are exactly what ATLAS 2026 describes as the new attack surface. Having skills that map to those specific threats rather than to generic “AI security” bullet points changes the conversation from abstract to actionable.
D3FEND is MITRE’s defensive complement to ATT&CK. Where ATT&CK catalogues what attackers do, D3FEND catalogues what defenders can do in response — specific countermeasures, detection methods, and hardening techniques mapped directly against the corresponding offensive techniques. Skills mapped to both ATT&CK and D3FEND give you attack context and mitigation in the same breath, which is how real security practitioners actually think about a technique.
NIST AI RMF — the AI Risk Management Framework — matters in 2026 specifically because Colorado’s AI Act (effective February 2026) provides a legal safe harbor for organizations that comply with it. If you’re building products in a regulated industry, having your security work traceable to AI RMF controls is no longer a nice-to-have. It’s the kind of documentation a compliance audit actually asks for.
MITRE F3 (Fight Fraud) extends the family into financial crime: social engineering, fraud detection patterns, identity verification failures, account takeover techniques. Less relevant for pure infrastructure security, directly relevant for anyone building fintech, e-commerce, or anything touching payment flows.
The reason mapping to all six simultaneously matters: real security work doesn’t live cleanly in one framework. A penetration tester needs ATT&CK. A CISO doing risk planning needs NIST CSF. An engineering team building AI products needs ATLAS and AI RMF. A fintech team needs F3. A library that forces you to pick a lane means you’re always pulling from multiple sources and doing the cross-mapping yourself. This one does it for you.
29 Security Domains — What 817 Skills Actually Covers 🔍
“817 sounds like a big number until you see what they’re actually covering and realize it might still be too few.”
The library organizes its 817 skills across 29 security domains. I want to call out a few that specifically caught my attention as a mobile developer who thinks about AI tooling:
Web security covers SQL injection, XSS, CSRF, SSRF, XXE, authentication bypass, session management vulnerabilities, and API security — at technique level, not just “be aware these exist” level.
Mobile security is the one I went deepest on. Deep link hijacking, Android intent abuse, iOS URL scheme vulnerabilities, insecure data storage patterns, certificate pinning bypass techniques. These are the exact attack vectors that matter for Flutter apps shipping to both platforms, and seeing them treated as first-class security domains rather than afterthoughts felt like finding a textbook written specifically for what I work on.
AI/ML security covers adversarial examples, model extraction attacks, prompt injection against LLMs, training data poisoning, and model inversion — the attack surface that most of us are actively building into products right now, often without the vocabulary to reason about the risks precisely.
Malware analysis — given my recent piece on the pub.dev XCSSET malware discovery, where a Flutter package was found carrying a macOS worm embedded in its build files, having structured skills for static and dynamic malware analysis feels immediately relevant rather than theoretical.
Reverse engineering includes something I found genuinely remarkable: MCP integration with Ghidra, the NSA’s open-source reverse engineering framework. This means an AI agent can decompile and analyze a binary directly using real tooling, not just reasoning about it from a description. The gap between “the agent thinks about the binary” and “the agent runs Ghidra against the binary” is enormous in practice.
DFIR (Digital Forensics and Incident Response) covers what you need when something’s already gone wrong — evidence collection, log analysis, memory forensics, timeline reconstruction. Most developers never touch this domain until they’re in the middle of an incident and wishing they’d understood it earlier.
Cloud security, threat intelligence, network security, cryptography, red team operations, OSINT, DevSecOps, compliance — all present, all mapped to the frameworks above.
The Progressive Disclosure Pattern — The Design Choice That Actually Makes This Usable 🏗️
“Most security resources are written for one audience. This one figured out how to serve three at once.”
Here’s the architectural detail I want to highlight specifically, because it separates this from a library that looks impressive on a README but frustrates you in practice. Each skill uses what Mukul calls a Progressive Disclosure Pattern — skills are layered so that a developer new to security gets enough context to understand and start, an intermediate practitioner gets methodology and proper technique depth, and an experienced security engineer gets full technical detail and direct tool invocations. All from the same skill. Invoked the same way.
This matters enormously because security knowledge has a steep and unforgiving gradient. The space between “I’ve heard of SQL injection” and “I can identify, exploit, and remediate this specific injection point in this specific codebase” is huge, and most resources live firmly on one end or the other. Skills that calibrate to where you actually are — and pull you forward rather than leaving you stranded — are the ones that get used across a career rather than installed once, skimmed once, and forgotten.
The Growth Story — 27,000 Stars in Three Months 📈
“Numbers that move that fast usually mean something genuinely resonated, not just that the marketing was good.”
The repo appeared on GitHub Trending on May 25, 2026, picking up over 999 new stars in a single 24-hour window. By the time I found it, it had passed 27,000 stars and 3,200+ forks. It was featured at Black Hat USA 2026’s Arsenal Lab — a juried selection process, not a sponsored appearance. It’s been covered independently across security newsletters, AI tooling blogs, and developer communities with no obvious coordinated push behind any of it.
That kind of organic spread for an open-source skill library in roughly three months says something specific: there was genuine unmet demand for exactly this. The skills ecosystem around AI coding agents has moved fast on productivity tooling — the security side was a real gap, and this filled it with more rigor and framework coverage than most people expected from a community project.
How to Actually Install It ⚡
“Fifteen seconds. Here’s exactly what to run.”
# Inside Claude Code — the cleanest path
/plugin marketplace add mukul975/Anthropic-Cybersecurity-Skills
# Via npx if you prefer terminal
npx skills add mukul975/Anthropic-Cybersecurity-Skills
# Manual install — works with any agent that reads structured Markdown
git clone https://github.com/mukul975/Anthropic-Cybersecurity-Skills.git ~/.claude/skills/cybersecurity
After install, skills live as plain Markdown files in ~/.claude/skills/ — completely readable and auditable before you invoke them. Given that my Claude Code setup article is literally about not trusting things that run in your agent context without review, I want to be clear: I read the skill files before using them, and they hold up. The structure is transparent. You can see exactly what instruction set you're loading before it runs.
Invoke any skill by name from Claude Code — /security-review, /web-app-pentest, /dfir-triage — and the relevant playbook loads and executes in context. No API keys, no configuration, nothing to maintain.
What This Did to My Portfolio — And My Thinking 💭
The portfolio got its security audit section eventually. I used the web security and OWASP skills from this library and it was faster and more thorough than anything I’d have prompted from scratch — the agent ran through the checklist with framework-level rigor instead of the vague “check for common vulnerabilities” output you get when you prompt without structured context.
But the more lasting thing this changed was how I think about the skills layer of AI-assisted development in general. I’ve been building the context layer carefully — CLAUDE.md, hooks, memory, all the stuff in my setup article — and mostly treating security as something I think about when I think about it. Finding a library that treats security knowledge as something you can structure, install, and invoke like any other skill shifts the question from “do I remember to think about security” to “does the skill fire at the right moment.”
That’s a different kind of reliability, and it’s the kind worth building on.
Tags: Claude Code, Cybersecurity, AI Coding Agents, MITRE ATT&CK, Developer Security, Open Source, Developer Tools
Before you go
- Please take a moment to like the post and follow the writer!
- Did you know that over 400,000 developers share what they’re building, learning, and discovering across our platforms every month? Learn how you can contribute here.
The Claude Code Security Library That Made Me Forget About My Portfolio was originally published in Stackademic on Medium, where people are continuing the conversation by highlighting and responding to this story.