Tools

Anthropic-Cybersecurity-Skills: 817 Security Skills Packaged for AI Agents

A 30k-star open-source repo splits six major security frameworks (MITRE ATT&CK, NIST CSF, ATLAS, and more) into 817 standardized SKILL.md files. Load them into Claude Code, Cursor, Copilot, or Codex CLI with a single config line.

Font size

What problem it solves

A junior security analyst knows which Volatility 3 plugin to use on a memory image and which Sigma rule catches Kerberoasting. A default AI agent knows neither.

This repo takes 817 structured security skills—each one a SKILL.md—and organizes them across 29 security domains and six major frameworks. Any tool that follows the agentskills.io standard can load and invoke them like native skills.

Think of it as the Blue Team Manual and the Red Team Manual, shredded into SKILL.md files an LLM can actually read.

How to use it

Three steps:

  1. git clone https://github.com/mukul975/Anthropic-Cybersecurity-Skills
  2. Add the repo path to the skills directory of Claude Code, Cursor, or Codex CLI:
    • Claude Code: usually ~/.claude/skills/
    • Cursor: .cursor/skills/ inside the project
  3. In the chat, just say something like "use the memory-forensics-volatility skill to analyze this dump."

Here is a minimal runnable skill file, pulled from the repo structure:

---
name: memory-forensics-volatility
description: |
  Analyze Windows memory dumps using Volatility 3 to identify malicious
  processes, injected code, and credential artifacts. Use when the user
  provides a .raw or .dmp file and asks about memory forensics, process
  injection, or credential theft (T1003 OS Credential Dumping).
framework_mapping:
  attack:
    - T1003
  csf:
    - DE.CM-01
---

# Memory Forensics with Volatility 3

## Prerequisites
pip3 install volatility3

## Quick Analysis Workflow
1. Identify image profile:
   `vol -f memdump.raw windows.info`
2. List processes (look for unsigned / hollowed):
   `vol -f memdump.raw windows.pslist`
3. Scan for code injection:
   `vol -f memdump.raw windows.malfind`
4. Dump credentials:
   `vol -f memdump.raw windows.hashdump`

## Output
Save findings to `findings.md` with MITRE ATT&CK technique IDs.

Drop this file into skills/memory-forensics/SKILL.md and Claude Code or Cursor will load it automatically when the right question comes up.

How it works

The core design rides on the agentskills.io open standard. Each skill lives in its own folder containing a SKILL.md: YAML frontmatter describes the skill name, trigger conditions, dependencies, and framework mappings; the body holds the actual steps and reference commands.

The 817 entries are sorted into 29 domain directories: forensics/, cloud/, ai-security/, fraud/, red-team/, and so on. Each domain has a README at its root explaining which frameworks it applies to.

Load flow: user asks a question → the agent retrieves matching SKILL.md files → loads them into the context window → runs the commands inside → returns results. Matching relies on semantic search over the description field.

Framework mapping works both ways. The Volatility example above carries ATT&CK's T1003 (OS Credential Dumping) and NIST CSF's DE.CM-01 (Continuous Monitoring) at the same time. One skill, viewed through several frameworks, mapped to the same block of work—handy when you're writing a compliance report.

The 29 domains cover attack surface management, threat intel, detection engineering, IR, digital forensics, cloud security, AI/ML security, fraud detection, red team, C2, phishing, ICS/OT, mobile, compliance, and privacy. Basically the full workflow of a mid-size SOC.

Compatibility with 26+ AI platforms is the headline feature. Claude Code, GitHub Copilot, Cursor, OpenAI Codex CLI, Gemini CLI all consume the same standard. Write one SKILL.md, run it everywhere.

What's good / what's not

Good:

  • One standard, many platforms. Switch tools without rewriting skills—migration cost is near zero.
  • Comprehensive framework mapping. Goes well beyond a single Sigma ruleset, with ATT&CK / NIST CSF / D3FEND / ATLAS / AI RMF / F3 codes attached.
  • Apache 2.0 license. Commercial use is fine, internal distribution is clean.
  • The README draws clear legal lines: dual-use skills like red-team C2, phishing simulation, and exploitation are marked "only on systems you're authorized to test." More thoughtful than most.

Not so good:

  • The 30k stars are partly a halo effect from "Anthropic" in the repo name. The README disclaimer makes clear the author is mukul975 and this is a community project, with no official connection to Anthropic PBC. Worth keeping in mind.
  • SKILL.md is prompt text for an LLM, not executable code. It's a knowledge base, not an automation engine. Complex scenarios still need a human writing code.
  • Some red-team and phishing skills carry real risk. Enterprises in regulated industries should run legal and security review before adoption.
  • Maintenance depends on a single maintainer. Bus factor of 1. Long-term sustainability is an open question.

I tried it

After cloning locally, I pointed Cursor's .cursor/skills/ at the skills/ directory and tested the Volatility skill under memory-forensics.

Prompt: "Analyze this memdump.raw and find suspicious processes." The agent loaded the skill correctly, produced the standard command flow volatility3 -f memdump.raw windows.pslist --pid, and flagged that pip3 install volatility3 was needed first.

From prompt to complete suggestion, latency was under 2 seconds. That's pure retrieval plus loading SKILL.md into context—not including the actual Volatility run. Volatility3 on a large memory dump still takes minutes to tens of minutes.

Switching to Claude Code produced the same results.

I also tried a Sigma rule skill for network-layer detection. The agent translated it into a valid Splunk SPL query, saving me from digging through Sigma docs.

Bottom line: treat it as a security manual for AI. Don't expect it to make automation decisions for you. Running commands, writing PoCs, kicking off scans—that's still human work.

Who should use it

A good fit for:

  • Blue team / SOC analysts who want to pull up detection rules while running Claude or Cursor in their IDE. SKILL.md beats digging through docs every time.
  • GRC / compliance roles who need to cite evidence against MITRE ATT&CK or NIST CSF. The framework mapping saves real time.
  • Security researchers learning the six major frameworks systematically. 29 domains of structured material is a solid starting point.
  • AI security practitioners focused on ATLAS and AI RMF. Ready-made skill templates are here.

Not a fit for:

  • Anyone hunting a turnkey automated scanning tool. This is a knowledge base, not an engine.
  • Red teams looking to run unauthorized pentests. The legal risk is yours, and the repo explicitly forbids it.
  • Total newcomers to security. Skills without judgment are dangerous—you can't blindly trust AI output, and running the wrong command from a SKILL.md can wreck things fast.

Do this / don't do that

Concrete suggestions:

  1. Clone locally first. Don't trust a cloud version number.
  2. Roll out gradually: get 1–2 skills working on your workstation (e.g., memory-forensics plus one Sigma rule) and verify that the agent loads them and generates commands as expected.
  3. Write an "authorized scope" doc for the team. Which skills apply to which systems, who signs off, where the red lines are. Pin it in your wiki.
  4. Human-review critical commands before they run. SKILL.md is a prompt, not a sandbox. Whatever shell access the agent has, you gave it.

Don't:

  • Don't mount the entire skills/ directory on a production environment. Loading 817 skills at once blows out the context window and floods retrieval with noise.
  • Don't run red-team skills against production "for security testing" without going through change management and getting written authorization.
  • Don't treat it as a SOAR or automation platform. It's a knowledge base. Orchestration still needs a human.

Sources

Source: https://github.com/mukul975/Anthropic-Cybersecurity-Skills