I mapped 754 cybersecurity skills to 5 frameworks so your AI agent doesn't have to wing it

AI agents are everywhere in 2026. They write code, triage alerts,
analyze logs, scan infrastructure. But ask one to investigate a
suspicious memory dump or hunt for C2 beaconing and it improvises.
No structure. No framework alignment. No verification steps.

That's the gap I've been working on.

What I built

Anthropic Cybersecurity Skills is an open-source library of 754
structured cybersecurity skills for AI agents. Every skill is a
self-contained directory:
skills/performing-memory-forensics-with-volatility3/
├── SKILL.md ← YAML frontmatter + step-by-step workflow
├── references/
│ ├── standards.md ← framework mappings
│ └── workflows.md ← deep technical procedures
├── scripts/
│ └── process.py ← functional helper scripts
└── assets/
└── template.md ← report templates

Each SKILL.md has YAML frontmatter for agent discovery and a
structured Markdown body for execution. The design is built around
progressive disclosure — irrelevant skills cost ~30 tokens to scan,
relevant ones provide complete expert-level guidance.

v1.2.0 — the five-framework release

Today I shipped the update I've been working toward since launch.
754 skills now mapped to 5 industry frameworks simultaneously.

Framework	Skills mapped	What it covers
MITRE ATT&CK Enterprise	754 / 754	Adversary tactics and techniques
NIST CSF 2.0	754 / 754	Cybersecurity risk management
MITRE ATLAS v5.5	81	AI/ML adversarial threats
MITRE D3FEND v1.3	139	Defensive countermeasures
NIST AI RMF 1.0	85	AI risk management

No other open-source library does this.

Why five frameworks?

Each one serves a different audience and a different question.

ATT&CK answers: what technique is the adversary using?

NIST CSF 2.0 answers: which risk management function does
this skill address? (Identify, Protect, Detect, Respond, Recover,
or the new Govern function)

MITRE ATLAS answers: if the target is an AI or ML system,
which adversarial technique applies? Model poisoning, prompt
injection, supply chain compromise, escape-to-host from an
agentic container — these have no ATT&CK equivalents. ATLAS
v5.5 added agentic AI techniques in the last two releases.

D3FEND answers: what do you actually DO to defend against it?
ATT&CK maps attacks. D3FEND maps the 267 countermeasures that
stop them. A skill like detecting suspicious PowerShell execution
now tells your agent: this counters T1059.001, and here are the
D3FEND defensive techniques (D3-EWF, D3-PSA) that apply.

NIST AI RMF answers: where does this fit in the AI risk
lifecycle? With the EU AI Act's full requirements going live
August 2 and Colorado's AI Act citing NIST AI RMF as legal
safe harbor, this mapping matters right now.

What the frontmatter looks like

name: detecting-prompt-injection-attacks
description: >-
  Detect and prevent prompt injection attacks against LLM
  applications, AI agents, and chatbot interfaces. Covers
  direct injection, indirect injection via retrieved content,
  jailbreak detection, and input validation strategies.
domain: cybersecurity
subdomain: ai-security
tags: [prompt-injection, ai-security, llm, T1059.001]
frameworks:
  mitre-attack: [T1059.001, T1078]
  nist-csf: [DE.CM-01, DE.AE-02]
  mitre-atlas: [AML.T0017, AML.T0051]
  mitre-d3fend: [D3-IDA, D3-ODA]
  nist-ai-rmf: [MEASURE-2.7, GOVERN-6.1]

Five framework fields. One skill. Zero manual mapping required.

What's in the 754 skills

26 security domains. The top ones by skill count:

• Cloud Security (60) — AWS S3 audits, Azure AD review, GCP IAM
• Threat Hunting (55) — C2 beaconing, DNS tunneling, LOTL detection
• Threat Intelligence (50) — APT attribution, campaign analysis, IOC enrichment
• Web App Security (42) — HTTP smuggling, XSS, deserialization
• Network Security (40) — Wireshark analysis, Suricata tuning, VLAN segmentation
• Malware Analysis (39) — Ghidra, YARA, .NET decompilation
• Digital Forensics (37) — Volatility3, disk imaging, browser artifacts

Plus OT/ICS, container security, zero trust, API security,
DevSecOps, mobile, cryptography, red teaming, and more.

How agents actually use this

Your agent scans frontmatters first (~30 tokens each). When a
skill matches the task, it loads the full SKILL.md and references.
Here's what happens when a user says "check this memory dump for
credential theft":

1. Agent scans 754 frontmatters → finds 12 relevant skills
2. Loads top matches including performing-memory-forensics-with-volatility3
3. Follows the structured Volatility3 workflow
4. Maps findings to ATT&CK T1003 (Credential Dumping)
5. References D3FEND D3-PSMD for defensive recommendations
6. Outputs structured findings with framework references

No improvisation. No hallucinated tool flags. Structured output
with framework alignment baked in.

Install

npx skills add mukul975/Anthropic-Cybersecurity-Skills

Works with Claude Code, GitHub Copilot, OpenAI Codex CLI, Cursor,
Gemini CLI, and any MCP-compatible agent.

Contributing

Apache 2.0. PRs reviewed within 48 hours. The easiest first
contribution is adding MITRE ATT&CK technique IDs to the 74
incident-response skills that still need mapping — see Issue #1.

---

The repo hit 4,100 stars in a few weeks entirely from community
sharing. If this solves a problem you've been working around,
a star helps others find it.

github.com/mukul975/Anthropic-Cybersecurity-Skills