Anthropic Cybersecurity Skills for giving AI agents structured analyst workflows

A community-maintained library of 818 structured security skills that turns a generic LLM agent into a guided analyst, mapped to MITRE ATT&CK, NIST CSF, and four other frameworks for authorized research and defense.

Tool	mukul975/Anthropic-Cybersecurity-Skills — 818 structured cybersecurity skills for AI agents across 34 domains, Apache-2.0 licensed, ~32.8k stars
Category	AI agent knowledge base / prompt-engineering skill library (Python)
Primary Use	Loading expert analyst workflows into agent platforms like Claude Code, Copilot, and Gemini CLI so they follow structured, framework-mapped procedures during authorized assessments, hunting, and DFIR
Safe Use	The README states explicitly that offensive and dual-use skills are intended for authorized penetration testing, security research, defense, and education, against systems you own or have explicit written permission to test
Telemetry Note	The library itself is passive — it adds no binaries or agents to a target. Defenders should instead watch for the artifacts the skills describe (e.g. Sigma rules for Kerberoasting, EVTX anomalies), and monitor which skill directories an internal AI agent loads

Most security repositories on GitHub fall into a familiar shape: wordlists, payloads, one-off scanners. mukul975/Anthropic-Cybersecurity-Skills takes a different angle entirely — it is not code that runs against systems, but a structured knowledge base that teaches an AI agent how a senior analyst thinks. The repository packages 818 Markdown-based skills across 34 security domains, each written against the agentskills.io open standard so that agent platforms can discover and load them generically. Despite the name, the README is careful to note this is an independent community project, not affiliated with Anthropic.

The architecture is deliberately boring in the best way. Each skill is a directory with YAML frontmatter for fast, sub-second discovery and a Markdown body encoding a step-by-step practitioner workflow, plus reference files for deep context. That means an agent host such as Claude Code, GitHub Copilot, OpenAI Codex CLI, Cursor, or Gemini CLI can filter skills by metadata — domain, technique ID, framework mapping — before ever reading the body. The claimed 26+ platform compatibility comes from adherence to the standard rather than per-platform forks, which is the right design decision for something intended to be widely consumed.

What makes the library genuinely interesting to a professional audience is the framework mapping discipline. Every skill carries frontmatter references to the industry frameworks relevant to its subject: MITRE ATT&CK v19.1 coverage spans 805 of the skills, NIST CSF 2.0 covers 804, while MITRE D3FEND, NIST AI RMF, MITRE ATLAS, and the newer MITRE F3 apply selectively. The README is explicit that not every skill hits every framework — a forensics skill carries ATT&CK plus CSF, while an AI-security skill adds ATLAS and AI RMF. That per-skill selectivity is more honest than blanket mapping, and the ATT&CK IDs were validated with the official mitreattack-python library against v19.1, with the authors claiming zero revoked or deprecated IDs.

The tactical distribution table is worth reading closely because it tells you where the library's weight sits. Initial Access (TA0001) covers 467 skills, Privilege Escalation 464, Persistence 444 — the offensive middle of the kill chain is heavily populated, which reflects the repo's topics (red-team, penetration-testing, ethical-hacking). But the defensive side is not an afterthought: 139 skills map to D3FEND countermeasures, and domains like SOC Operations (63 skills), Threat Hunting (58), and Digital Forensics (41) give blue teams substantial material. For a defender, the same mappings work as an index into detection engineering — pick an ATT&CK technique and see what workflows exist around it.

The newest addition is mapping against MITRE F3, the Fight Fraud Framework released in April 2026 by MITRE's Center for Threat-Informed Defense. F3 fills the post-compromise gap ATT&CK leaves open — what an adversary does after access to actually monetize it — with tactics like Positioning (FA0001) and Monetization (FA0002), and technique IDs in the F1XXX namespace. 94 skills carry F3 mappings, and the authors state all 123 F3 v1.1 technique IDs were verified against the upstream STIX bundle. For financial-sector defenders this is the most differentiated part of the library; detecting-business-email-compromise, for example, maps to F1005.006 under monetization alongside the classic T1566.

Domain coverage reads like a competency matrix for a full-spectrum team. Cloud Security leads with 66 skills spanning AWS, Azure, and GCP hardening plus cloud forensics; Threat Intelligence offers 52 skills around STIX/TAXII, MISP, and OpenCTI; Malware Analysis contributes 39 covering static/dynamic analysis and sandboxing; Red Teaming has 35 including ADCS abuse with Certipy, path analysis with BloodHound CE, and C2 tooling. The identity domain mixes offensive and defensive material — Entra ID attack paths via ROADtools sit next to PAM and zero-trust identity design — which is consistent with the repo's stated dual-use, authorized-only posture.

It is worth pausing on the framing the README itself provides, because it is unusually direct for a repo with red-team content. A prominent disclaimer states the library includes offensive and dual-use techniques — explicitly naming C2, phishing simulation, and exploitation — and restricts them to authorized penetration testing, security research, defense, and education, against systems you own or have written permission to test. There is nothing in the repository that executes autonomously; the skills are procedural knowledge consumed by an agent a human operates. From a documentary standpoint, this is knowledge curation, not weaponization, and it should be evaluated the way you would evaluate an ATT&CK-mapped internal wiki.

Installation is a one-liner either way: npx skills add mukul975/Anthropic-Cybersecurity-Skills for the recommended path, or a plain git clone https://github.com/mukul975/Anthropic-Cybersecurity-Skills.git if you want to audit the content before your agent ever sees it. The audit path deserves emphasis — if you are piping 818 skills into an agent that executes tool calls in your environment, you are effectively installing a large third-party policy corpus. Reviewing the SECURITY.md and CONTRIBUTING.md first, and loading only the domains you need, is sensible supply-chain hygiene for AI agent deployments.

There are caveats a senior operator should hold onto. The README interleaves the core content with promotional material — a survey badge, a Casky.ai playground waitlist promising live exercises — which does not affect the skill corpus itself but does mean the document is part documentation, part marketing funnel. Treat the external service claims as unverified and evaluate the repository on its local artifacts. Similarly, the 4.8-million workforce-gap statistic from ISC2 is context for motivation, not a technical claim you should lean on in a risk memo.

Where this fits in an authorized workflow is straightforward: it is a force multiplier for junior staff and agent-assisted triage. A first responder asking an agent which Volatility3 plugin suits a suspicious memory dump, or which Sigma rules catch Kerberoasting, gets a structured answer grounded in mapped, versioned technique IDs instead of generic model output. Teams running tabletop exercises, hypothesis-driven hunts, or cloud-posture reviews can pull the relevant domain and get a repeatable checklist with references — and because everything is plain Markdown under Apache-2.0, forking and internalizing the library into your own knowledge base is fully permitted.

As a defensive artifact, the library also functions as a mirror: whatever skills exist for red-team domains, the corresponding detection knowledge exists in the hunting, SOC, and D3FEND-mapped entries. Reading both sides of a technique — the offensive workflow and its detection counterpart — is precisely how detection engineering gets done, and having them co-located in one indexed corpus is the practical value proposition here. For anyone building agentic security tooling in 2026, Anthropic-Cybersecurity-Skills is less a tool to deploy than a dataset to study, fork, and hold your agents accountable to.

Official project repository for mukul975/Anthropic-Cybersecurity-Skills.

Download Tool

Educational analysis for authorized security professionals. Use only in controlled, authorized environments.