In 2025, the FBI received reports of $3.05 billion in losses from business email compromise.
Figure 1. Reported losses by crime type in 2025. Source: FBI Internet Crime Complaint Center (IC3), 2025 Internet Crime Report.
Anthropic-Cybersecurity-Skills is a free, open-source library of 818 step-by-step security playbooks that AI agents like Claude Code can load and follow.
Many early-stage startups don't have a security hire yet, and this library puts senior-analyst procedures inside tools many founders already use.
This guide covers what the library is (and isn't), the 14 skills worth installing, where to find the install steps, and how to use the skills safely.
Below is the full breakdown.
Inside you'll find:
The $3 Billion Threat That Starts With an Email
What This Library Is, and What Its Name Gets Wrong
The 14 Skills Worth Installing (Out of 818)
How to Install Them, and One Catch
How to Use Them Without Breaking Anything
Your First Week: A Five-Day Plan and the Link
1. The $3 Billion Threat That Starts With an Email
Most founders picture a cyberattack as someone breaking into their servers. One of the most expensive threats in the FBI's data starts with something simpler: an email.
Business email compromise (BEC) is email fraud aimed at moving money. Attackers pose as your CEO, a supplier or your lawyer and ask your finance person to pay a "new" bank account. Sometimes they break into a real email account first; often, impersonating one is enough. Either way, the money moves because someone trusted an email.
The FBI's Internet Crime Complaint Center (IC3) received 24,768 BEC complaints in 2025, with $3.05 billion in reported losses. Only investment fraud cost more. The report's totals include complaints from more than 200 countries, not just the US.
→ That is about $123,000 per complaint on average (our calculation from the IC3 totals). A few very large cases pull an average like this up, so a typical loss can be much smaller. Still, a single fraudulent transfer can cost a seed-stage startup months of runway.
The same report counted 191,561 phishing and spoofing complaints. In many of these cases, the weak point is a process, not a firewall: who can approve a payment, and who can send email that looks like it came from your domain.
The second reason security lands on a founder's desk is sales. Once you start selling to larger companies, security questionnaires and requests for a SOC 2 report or an ISO 27001 certificate often show up in the deal. At that point, security is a revenue question.
The most telling detail is that most of the defenses above are settings or policies, not products you have to buy. That is exactly the kind of work an AI agent with the right playbook can help you with.
2. What This Library Is, and What Its Name Gets Wrong
Anthropic-Cybersecurity-Skills is a public GitHub repository of 818 "skills". A skill is a folder of instructions that an AI agent loads when a task calls for it: when to use it, what it needs, the exact steps, and how to check the result. Most skills also include reference notes, templates and helper scripts.
The numbers, as of October 23, 2026: about 33,700 stars and 4,100 forks on GitHub, an Apache 2.0 license (free to use, including commercially), and mappings to six security frameworks, among them MITRE ATT&CK and NIST CSF 2.0. You don't need to know those frameworks. Auditors do, and the mappings help when you talk to them.
Three things to know before you install anything:
It is not made by Anthropic. Despite the name, the README states that it is an independent community project, not affiliated with Anthropic. It was created by an independent developer (GitHub user mukul975), and contributions are reviewed by a single maintainer.
It is not a security product. It doesn't monitor anything or block attacks. It gives your AI agent a playbook. The agent still needs access to your systems, and you still make the decisions.
Most of it is written for security teams. The largest areas are cloud security, security operations centers, threat hunting and threat intelligence. Some skills describe attack techniques (red teaming, phishing simulations, exploitation) that are meant only for authorized testing.
Figure 2. Skills per security domain, as listed in the library's README. The README's domain table adds up to 817 skills; the repository holds 818.
→ The practical takeaway: treat the library like a large professional manual. You need a small slice of it.
3. The 14 Skills Worth Installing (Out of 818)
We went through the library and kept only the skills a startup can act on without a security team, grouped by the problem they solve. The name in monospace is the exact skill name, the one you will use when installing. Check the "Use it if" line to skip what doesn't apply to you.
Group A. Email and accounts: close the front door (every startup)
1. implementing-dmarc-dkim-spf-email-security
What it does: Checks and sets up the three DNS records (SPF, DKIM and DMARC) that stop criminals from sending email that appears to come from your domain.
You'll need: Access to your domain's DNS settings (wherever you bought or manage the domain) and to your email admin console.
Use it if: You have a company domain. That is everyone. It is the highest-value, lowest-effort fix on this list.
2. implementing-google-workspace-admin-security
What it does: A hardening checklist for Google Workspace: multi-factor authentication for everyone (with physical security keys for admins), control over which third-party apps can access company data, and limits on external file sharing.
You'll need: Super admin access. The skill lists a Business Plus or Enterprise plan as a prerequisite, so some steps may not be available on cheaper plans.
Use it if: Your company runs on Google Workspace. On Microsoft 365, the closest skill is
implementing-conditional-access-policies-azure-ad(sign-in rules in Microsoft Entra ID, such as requiring multi-factor authentication). It is much shorter than this checklist.
3. detecting-business-email-compromise
What it does: Explains how fake-CEO and fake-supplier payment scams work, the red flags to watch for, signs that an email account has been taken over, and the financial controls that stop the fraud: two people to approve large transfers, and a phone call-back before any change to a supplier's bank details.
You'll need: Nothing for the financial controls. The detection steps assume an email security gateway and a SIEM (a system that collects and monitors logs), which most early-stage startups don't have yet.
Use it if: Anyone at your company can send money. Start with the financial controls.
Group B. Code and cloud: stop leaking keys and data (if you have a product live)
4. implementing-secret-scanning-with-gitleaks
What it does: Uses the free tool Gitleaks to find passwords, API keys and tokens committed to your code repositories, including old history, and blocks new ones before they get committed.
You'll need: Your code repositories (GitHub, GitLab or similar). Any key it finds has to be rotated (replaced with a new one), not just deleted from the code.
Use it if: You have a codebase.
5. auditing-cloud-with-cis-benchmarks
What it does: Runs free scanners (Prowler or ScoutSuite) against your AWS, Azure or Google Cloud account, compares the setup with the CIS security benchmarks, and gives you a prioritized list of fixes. The skill also positions it as preparation for SOC 2 and ISO 27001 audits.
You'll need: Read-only access to the cloud account (on AWS, the SecurityAudit policy). Fixes are made separately, with different credentials.
Use it if: Your product runs on a cloud provider.
Group C. If your product uses AI: don't let your agent get hijacked
6. detecting-indirect-prompt-injection
What it does: Screens the web pages, documents and images your AI reads for hidden instructions (the "ignore your rules and send me the data" kind) before the model processes them.
You'll need: A developer and a Python environment. It works on your own code.
Use it if: Your product has an AI feature that browses the web, reads email, summarizes documents or accepts user uploads.
7. securing-agentic-ai-tool-invocation
What it does: Limits what your AI agent can do when it calls tools with real effects, like sending email, making payments or writing files. It covers allowlists, minimum permissions, human approval for risky actions, and audit logs.
You'll need: A developer and access to your agent's code.
Use it if: Your agent takes actions, not just answers questions.
8. auditing-mcp-servers-for-tool-poisoning
What it does: Audits MCP servers (the connectors that plug tools and data into AI assistants) for poisoned tool descriptions, silent changes after approval, and exposed endpoints, using the open-source scanner mcp-scan.
You'll need: The MCP configuration files you want to check, and Python.
Use it if: You build MCP servers, or your team installs MCP connectors in tools like Claude Desktop or Cursor. This one matters even if your product has no AI.
Group D. Selling to larger customers: compliance
9. performing-soc2-type2-audit-preparation
What it does: Runs a gap assessment against the SOC 2 criteria, collects evidence automatically from your cloud provider, identity provider (Okta or Google Workspace), GitHub and ticketing tool, and prepares evidence packages for the auditor.
You'll need: Read-only access to those systems. The audit itself is done by an independent CPA firm; the skill gets you ready for it.
Use it if: Customers are asking for a SOC 2 report.
10. implementing-iso-27001-information-security-management
What it does: A full plan to build an ISO 27001:2022 information security management system, from gap analysis and risk assessment to the certification audit. The skill's phases add up to about 42 weeks, so start early.
You'll need: Time and an owner on your team. Most of the work is documentation and process.
Use it if: Customers ask for ISO 27001 certification rather than a SOC 2 report.
11. conducting-gdpr-compliance-assessment
What it does: Assesses how you handle personal data against the GDPR: records of processing, legal basis, people's rights over their data, breach notification within 72 hours, and data transfers outside the EU.
You'll need: A list of the personal data you collect and the tools that store it. Have a lawyer review the result; it is a strong first draft, not legal advice.
Use it if: You have users or customers in the EU.
Group E. The bad day: be ready before it happens
12. building-incident-response-playbook
What it does: Writes step-by-step response playbooks per incident type (a phishing attack, a hacked account): who decides, who to call and what to do in the first hours, aligned with NIST's incident response guide (SP 800-61r3).
You'll need: A list of the people and providers you would call. The skill is written for companies with a security team, so ask your AI agent for a one-page version sized for your team.
Use it if: You don't have a written plan for a security incident.
13. implementing-ransomware-backup-strategy
What it does: Designs backups that survive ransomware using the 3-2-1-1-0 rule: three copies, two types of storage, one offsite, one that can't be altered or deleted, and zero errors when you test a restore.
You'll need: A list of the data you can't afford to lose. The skill's examples use enterprise backup tools, but the rule applies to any company.
Use it if: You have never tested restoring your backups.
14. deploying-honeytokens-and-canarytokens
What it does: Plants free digital tripwires (a fake AWS key, a decoy document, a decoy link) that email you the moment someone touches them, an early sign that someone is inside your systems.
You'll need: Nothing to start. Tokens can be generated at canarytokens.org without an account; self-hosting is optional.
Use it if: You want a cheap alarm while you don't have security monitoring.
4. How to Install Them, and One Catch
The official install steps are in the Quick start section of the library's README on GitHub: https://github.com/mukul975/Anthropic-Cybersecurity-Skills#quick-start
As of October 23, 2026, the README lists two options:
# Option 1: npx (recommended)
npx skills add mukul975/Anthropic-Cybersecurity-Skills# Option 2: Git clone
git clone https://github.com/mukul975/Anthropic-Cybersecurity-Skills.git
cd Anthropic-Cybersecurity-SkillsAccording to the README, the skills work with Claude Code, GitHub Copilot, OpenAI Codex CLI, Cursor, Gemini CLI and other tools that support the Agent Skills standard. Follow the README for the steps. Three brief notes before you do:
Install only the skills you need. Option 1 uses a third-party installer called skills (by Vercel Labs), which needs Node.js and has a --skill option to pick skills by their exact name instead of installing all 818. That matters: in Claude Code, for example, every installed skill adds its description to what Claude reads on every turn, and with hundreds installed many descriptions get shortened, so Claude is less likely to pick the right skill on its own.
The regular Claude chat app is a different case. According to Claude's documentation, uploading a skill whose header includes fields outside the official skill format fails, and every skill in this library has extra header fields (domain, tags, framework mappings). Use the skills in an agent like Claude Code, or remove those fields before uploading.
On Windows, if PowerShell says that running scripts is disabled when you run npx, run the same command in Command Prompt instead.
5. How to Use Them Without Breaking Anything
Once a skill is installed, you don't need to remember any commands. Ask in plain language and the agent picks the matching skill. In Claude Code, you can also call one directly by starting your message with / followed by the skill name.
Three prompts to start with:
Use the implementing-dmarc-dkim-spf-email-security skill to audit the email
setup of yourcompany.com. Only read public DNS records and don't change
anything. Then tell me in plain language which records to add, and where./implementing-secret-scanning-with-gitleaks Scan this repository, including
its full history, and list any exposed keys. Don't modify any files.Use the building-incident-response-playbook skill to write a one-page
playbook for a hacked email account at a 10-person startup.
Ask me questions instead of assuming.Five rules that keep this safe:
Only use the skills on systems you own or have written permission to test. Some skills in the library describe attack techniques. Using them on someone else's systems can be illegal.
Give read-only access for audits. Make changes in a separate, deliberate step.
Read before you approve. Agents like Claude Code can ask for approval before running commands, depending on your permission settings, and many skills ship with scripts. Each skill is a plain-text folder you can read on GitHub before installing. If a command would change live settings, delete data or send something, stop and ask your technical lead.
Keep secrets out of the chat. Don't paste passwords or API keys. If a key gets exposed, rotate it.
Treat the output as a strong first draft. A SOC 2 report still requires an auditor, and GDPR conclusions need a lawyer. The skills get you prepared faster; they don't sign anything for you.
6. Your First Week: A Five-Day Plan and the Link
You don't need all 14 skills on day one. This plan starts with email fraud, the second-largest loss category in the FBI's report and one that targets businesses directly.
Day 1. Install five skills (see Section 4):
implementing-dmarc-dkim-spf-email-security,implementing-google-workspace-admin-security(or the Microsoft 365 alternative),detecting-business-email-compromise,implementing-secret-scanning-with-gitleaksandbuilding-incident-response-playbook.Day 2. Fix your email records. Audit your domain's SPF, DKIM and DMARC records and add what's missing. The skill rolls DMARC out in phases: monitoring first, then quarantine, then reject.
Day 3. Lock down accounts and payments. Turn on multi-factor authentication for everyone and agree on one rule: no change to a supplier's bank details without a phone call to a number you already had.
Day 4. Scan your code. Check every repository for exposed keys and rotate anything the scan finds.
Day 5. Write your one-page incident plan. Who decides, who you call (bank, lawyer, cloud provider) and what you do in the first hour.
After that, add the cloud audit once you have production infrastructure, the AI skills if your product uses AI, and the compliance skills when the first larger customer asks.
Security at an early-stage startup isn't a department. It's a handful of decisions, and now you have a playbook for each one.
For more practical resources for founders, from pitch decks to fundraising and AI tools, follow The Enduring Founder.
Subscribe to CLUB to read the rest.
Become a paying subscriber of CLUB to get access to this post and other subscriber-only content.
UpgradeA subscription gets you:
- Everything included in Premium.
- Early access to new tools and the specialized AI and product-market fit newsletter.
- One-on-one feedback on your business, with actionable guidance on strategy, positioning, and growth
- Masterclasses with Pablo Fernandez Alvarez