754 cybersecurity skills for your agent, one command away

Ask a vanilla AI agent to "analyze this memory dump for credential theft," and it will produce a reasonable-sounding checklist: check running processes, scan for suspicious network connections, look for unusual registry entries. What it won't do is run windows.pslist against psscan to cross-view-detect hidden processes, then pull hashdump + lsadump to extract cached credentials, then map the findings to ATT&CK T1003 — because that requires knowing which Volatility3 plugins to chain in what order. That gap is what mukul975/Anthropic-Cybersecurity-Skills closes. 1

13,500 stars, 1,600 forks, Apache 2.0 — this is the first security-specific skills library in the Agent Skills ecosystem, and its architecture is worth understanding before you install it. 1

⚠️ The name is misleading: despite "Anthropic" appearing in the repo title, this is an independent community project by Mahipal Jangra (GitHub: mukul975) with no affiliation to Anthropic PBC. 1 The name refers to its use of the agentskills.io open standard, which Anthropic helped define.

github.com · GitHub リポジトリ

mukul975/Anthropic-Cybersecurity-Skills

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

コンテンツカードを読み込んでいます…

The architecture that makes 754 skills fit in one context window

Most skill libraries front-load everything into the agent context — which hits a wall fast when you have hundreds of skills. This repo solves that with progressive disclosure: each skill costs approximately 30 tokens to scan (YAML frontmatter only) and 500–2,000 tokens to fully load (complete workflow). 2

The agent's lookup flow:

Scan all 754 frontmatters in one pass (~22,600 tokens total)
Match tags/domain/description against the user's prompt → identifies ~12 candidates
Load the top 3 full workflows (~1,500–6,000 tokens)
Execute step-by-step

Fazal, writing a detailed review on Medium, described it: "Claude doesn't load every skill into memory all at once (which would overwhelm its context window). Instead, it scans available skills by reading only the YAML frontmatter." 3

Every SKILL.md follows a fixed four-section structure: When to Use (exact trigger conditions), Prerequisites (tools, permissions, environment), Workflow (numbered steps with actual commands), Verification (how to confirm the output is correct). That consistency is what makes the frontmatter-scan approach reliable — the agent knows exactly what to expect once it loads a full skill.

What's inside: 26 domains, 5 compliance frameworks

The 754 skills 1 cover the major security operations disciplines:

Domain	Skills	Examples
Cloud Security	60	AWS misconfiguration detection, GCP IAM audit
Threat Hunting	55	Hypothesis-based hunt workflows, IOC pivoting
Threat Intelligence	50	CTI collection, indicator enrichment
Web App Security	42	OWASP Top 10 testing, SSRF/XXE exploitation
Network Security	40	Packet analysis, protocol anomaly detection
Malware Analysis	39	Static/dynamic analysis, sandbox setup
Digital Forensics	37	Memory forensics, disk triage, chain of custody

Each skill maps to all five of these frameworks simultaneously: 2

MITRE ATT&CK v19.1 — 286 techniques across 15 enterprise tactics (all 754 skills mapped)
NIST CSF 2.0 — all 6 functions (Govern / Identify / Protect / Detect / Respond / Recover)
MITRE ATLAS v5.5 — 81 skills mapped to AI/ML adversarial techniques (model poisoning, prompt injection defense)
MITRE D3FEND v1.3 — 139 skills mapped to defensive countermeasures
NIST AI RMF 1.0 — 85 skills mapped to AI risk management categories

That five-framework mapping per skill is what distinguishes this from a general security cheat-sheet — every workflow comes with pre-attached compliance evidence. Colorado's AI Act (effective February 2026) provides legal safe harbor to organizations that follow NIST AI RMF; having framework mappings baked into every skill output matters in that context. 2

Install (three paths)

Recommended — one command, all platforms:

npx skills add mukul975/Anthropic-Cybersecurity-Skills

Works immediately with Claude Code, GitHub Copilot, Cursor, Windsurf, Cline, Aider, Continue, Roo Code, Amazon Q Developer, OpenAI Codex CLI, Gemini CLI, Devin, SWE-agent, LangChain, CrewAI, AutoGen, and any MCP-compatible agent — 26+ platforms total. 2

Git clone (if you need offline access or want to inspect skills before loading):

git clone https://github.com/mukul975/Anthropic-Cybersecurity-Skills.git
cd Anthropic-Cybersecurity-Skills

Browser sandbox: Casky.ai runs a Playground environment for trying skills without local installation. It's still on a waitlist as of June 2, but worth bookmarking. 1

No npm package, no pip distribution — the only install vectors are npx skills and git clone.

Walkthrough: memory forensics with Volatility3

The performing-memory-forensics-with-volatility3 skill shows exactly what "structured operational knowledge" means in practice. 4

Prerequisites the skill checks before proceeding: Python 3.7+, Volatility3 installed via pip install volatility3, a raw/ELF/crash-dump format memory image, symbol tables (ISF files) for the target OS version, and disk space equal to 2–3× the memory image size.

The seven-step workflow the agent executes:

Acquire and verify the memory image (format identification, hash validation)
Identify OS profile via windows.info / banners
Enumerate processes: pslist → pstree → psscan cross-view to detect hidden processes → malfind for injected code
Analyze network connections and registry artifacts
Extract credentials: hashdump, lsadump, cachedump, cmdline (checks for Mimikatz artifacts)
YARA malware scan: yarascan, modules, modscan, ssdt
Compile a structured report with MITRE ATT&CK mappings

The Verification section gives a concrete output template: it expects the agent to surface hidden PIDs (from the pslist vs. psscan delta), confirmed injected process count (from malfind), and extracted credential hashes with their ATT&CK technique tags (T1003, T1059). 4 Without that verification gate, the agent's output is just text — with it, you get a binary pass/fail on whether the forensic chain is complete.

The prompt to trigger this in Claude Code or Cursor:

Use the performing-memory-forensics-with-volatility3 skill to
analyze memory.raw. Focus on credential theft indicators and
lateral movement artifacts. Map all findings to MITRE ATT&CK.

Hasan Toor (437K followers on X), who spotlighted the repo when it launched, put the library's value simply: "A database of real, organized security skills that any AI agent can plug into and use. Not tutorials. Not blog posts." 5

コンテンツカードを読み込んでいます…

Honest caveats

No quantitative benchmarks exist. Every effectiveness claim in the README is qualitative. There is no "with skill vs. without skill" response time comparison, no accuracy measurement across any of the 754 workflows. The repo does not claim otherwise — but you should factor this into how much you trust the workflows until you run your own comparisons. 1

Some domains are thin. Deception Technology has 2 skills. Compliance & Governance has 5. Mobile Security has 12. The author's CONTRIBUTING.md explicitly flags these as needing community help. 1

Frontmatter format inconsistency. The Volatility3 skill (an older file) uses mitre_attack: directly in the frontmatter, while newer skills (post-v1.2.0) use atlas_techniques:, d3fend_techniques:, and nist_ai_rmf: fields. Agents relying on consistent frontmatter structure across all skills may need handling for both formats. 4

ATT&CK v19.1 coverage gap. MITRE released ATT&CK v19 on April 28, 2026, splitting Defense Evasion (TA0005) into Stealth and Defense Impairment as separate tactics. The README notes this will be addressed in a forthcoming release, but as of June 2, no new tagged version has shipped. 6 Workflows that currently map to Defense Evasion may reference outdated technique IDs.

Token costs at scale. A Reddit user experimenting with the library noted: "I've spent a good amount of tokens in just a couple hours, unsustainable for personal use, lol." 7 The full-scan frontmatter pass runs ~22,600 tokens before any workflow loads. For personal or hobby-tier API access, this adds up.

コンテンツカードを読み込んでいます…

Last tagged release: April 6, 2026. The main branch has continued receiving commits, but no new version has been tagged since v1.2.0 — roughly two months ago. 6

When NOT to use this

Your threat model involves offensive operations. The skills are structured as defensive and analytical playbooks — incident response, forensics, threat hunting, compliance. They are not exploit development or active attack chaining workflows. Using them for red-team automation beyond authorized penetration testing is outside their intended scope and Apache 2.0 terms don't add guardrails.

You need the Deception Technology or Governance domains specifically. With 2 and 5 skills respectively, those domains are placeholders at this point, not operational libraries.

Your agent framework expects all skills to use the same frontmatter schema. The v1.0/v1.1 frontmatter format (using mitre_attack:) and the v1.2 format (using framework-specific fields) coexist in the same repo. Any orchestration layer that parses frontmatter fields programmatically needs to handle both.

You want ATT&CK v19.1 compliance today. The Defense Evasion tactic split is not yet reflected in the skill mappings. Organizations with strict ATT&CK version requirements should wait for the next tagged release.

Cover: repository banner via mukul975/Anthropic-Cybersecurity-Skills (Apache 2.0)

754 cybersecurity skills for your agent, one command away

The architecture that makes 754 skills fit in one context window

What's inside: 26 domains, 5 compliance frameworks

Install (three paths)

Walkthrough: memory forensics with Volatility3

Honest caveats

When NOT to use this

参考ソース