OWASP/www-project-agentic-skills-top-10

OWASP Agentic Skills Top 10

[]() []()

Security Risks and Mitigations for AI Agent Skills

> Covering OpenClaw (SKILL.md YAML), Claude Code (skill.json), Cursor/Codex (manifest.json), and VS Code (package.json) ecosystems.

Quick Links

• 🌐 Website: https://owasp.github.io/www-project-agentic-skills-top-10/
• 📖 Documentation: Full Documentation
• Tutorial Videos: Play the AST10 tutorial videos
• Public Review v1 Google Doc: Submit comments in Google Docs
• 🐙 GitHub Repository: https://github.com/OWASP/www-project-agentic-skills-top-10
• 🏛️ OWASP Project Page: https://owasp.org/www-project-agentic-skills-top-10
• 📋 Security Checklist: checklist.md
• 🔧 Universal Skill Format: universal-skill-format.md
• 📊 Case Studies: case-studies.md
• 🛡️ Threat Intelligence: threat-intelligence.md
• 🔍 Risk Assessment Tool: risk-assessment.md
• 🔎 Skill Scanner Integration: skill-scanner-integration.md
• 🔌 API Documentation: api-documentation.md
• 📚 Skill Development Guide: skill-development-guide.md
• ⚖️ Platform Comparison: platform-comparison.md
• 👥 Community & Contribution: community-contribution.md
• 🚨 Incident Response Playbook: incident-response.md
• 📊 Security Metrics & Monitoring: metrics-monitoring.md
• 🤝 Contributing: CONTRIBUTING.md
• 🛠️ Maintenance: MAINTENANCE.md

For the v1 publication cycle, please submit document review comments in the Google Doc above. Earlier comments, issues, and pull requests remain available in GitHub for transparency and historical reference.

---

Overview

The OWASP Agentic Skills Top 10 (AST10) is the first comprehensive security framework for AI agent skills. It documents the 10 most critical security risks in agentic AI skills across all major AI agent platforms, providing evidence-based mitigations and prevention strategies.

Key Features

• 🎯 Platform Coverage: OpenClaw, Claude Code, Cursor/Codex, VS Code
• 🔍 Risk Analysis: 10 critical security risks with real-world evidence
• 🛡️ Mitigation Strategies: Practical prevention and response guidance
• 📊 MAESTRO Mapping: Alignment with CSA's 7-layer threat model
• 🔗 Cross-References: Interconnected risk analysis
• 📝 Universal Format: Cross-platform skill security standard
• 👥 Community Driven: Open source with active contributor community

Why Agent Skills Matter

While significant attention has been given to securing LLMs and MCP tools, agent skills represent the behavior layer that translates AI capabilities into real-world actions. Skills define how agents orchestrate multi-step workflows, making them a critical attack surface.

2026 Statistics:

• 3,984 skills scanned across registries
• 36.82% contained security flaws
• 13.4% had critical vulnerabilities
• 76+ confirmed malicious payloads
• 1,184 skills in ClawHavoc campaign

---

Table of Contents

• Overview
• The Problem
• Risk Summary
• Getting Started
• Contributing
• Project Status
• Resources
• License

---

The Problem

This is not a theoretical future risk. The AI agent skill ecosystem is under active attack as of Q1 2026.

Active Threats (2026)

| Threat | Impact | Evidence | |--------|--------|----------| | ClawHavoc Campaign | 1,184 malicious skills | Antiy CERT (Feb 2026) | | Registry Poisoning | 36.82% skills vulnerable | Snyk ToxicSkills (Feb 2026) | | Claude Code RCE | CVE-2025-59536/21852 | Check Point Research (Feb 2026) | | WebSocket Hijacking | CVE-2026-28363 | Oasis Security (Feb 2026) | | Supply Chain Attacks | 280+ leaky skills | Snyk (Feb 2026) |

The Lethal Trifecta

An AI agent skill is especially dangerous when it has: 1. Access to private data (SSH keys, API credentials, wallet files) 2. Exposure to untrusted content (skill instructions, memory files, email) 3. Ability to communicate externally (network egress, webhook calls)

Most production agent deployments satisfy all three conditions.

---

Risk Summary

| # | Risk | Severity | Key Mitigation | |---|------|----------|----------------| | AST01 | Malicious Skills | Critical | Cryptographic signing, behavioral scanning | | AST02 | Supply Chain Compromise | Critical | Transparency logs, dependency pinning | | AST03 | Over-Privileged Skills | High | Least-privilege manifests, runtime enforcement | | AST04 | Insecure Metadata | High | Schema validation, safe parsers, sandboxed loading | | AST05 | Untrusted External Instructions | High | Source inventory, content pinning and inlining, continuous rescanning | | AST06 | Weak Isolation | High | Containerization, process isolation | | AST07 | Update Drift | Medium | Immutable pinning, hash verification | | AST08 | Poor Scanning | Medium | Multi-tool pipeline, semantic analysis | | AST09 | No Governance | Medium | Skill inventories, audit logging | | AST10 | Cross-Platform Reuse | Medium | Universal format, platform validation |

See index.md for detailed descriptions and MAESTRO mappings.

---

Getting Started

For Security Teams

1. Assess Current Posture: Use the Security Assessment Checklist 2. Review Risk Details: Read the 10 AST files for specific platform guidance 3. Implement Controls: Apply mitigations appropriate to your environment 4. Monitor Threats: Subscribe to security advisories and threat intelligence

For Skill Developers

1. Follow Best Practices: Implement least-privilege and secure coding 2. Use Universal Format: Adopt the proposed cross-platform standard 3. Sign Your Skills: Enable cryptographic verification 4. Test Security: Validate in isolated environments

For Platform Developers

1. Registry Security: Implement scanning and provenance tracking 2. Runtime Isolation: Default to sandboxed execution 3. Audit Logging: Enable comprehensive activity monitoring 4. User Trust: Require explicit confirmation for privileged operations

---

Contributing

We welcome contributions from the community! Here's how to get involved:

Ways to Contribute

• 🐛 Report Issues: Found a security risk or documentation error?
• ✨ Add Content: New examples, mitigations, or research
• 🔧 Code Examples: Provide secure coding patterns
• 🌐 Translations: Help localize the guide
• 📊 Research: Share threat intelligence or analysis

Getting Started

1. Read Guidelines: Check CONTRIBUTING.md 2. Fork & Clone: Work on your own branch 3. Use Templates: Issue and PR templates provided 4. Test Changes: Run validation scripts 5. Submit PR: Follow the contribution workflow

Submit New Risk Entries

Use our interactive web form to submit new AST risk entries:

➕ Submit New Risk Entry

The form generates properly formatted markdown and provides multiple submission options:

• Direct GitHub file creation
• Create GitHub Issue
• Download and manual PR

Community

• Mailing List: [email protected]

---

Project Status

Status: Active Development → Launch Preparation Version: 1.0 (2026 Edition) License: CC BY-SA 4.0

Timeline

| Phase | Status | Deliverables | |-------|--------|-------------| | Foundation | ✅ Complete | Repo, OWASP page, AST01-10 drafts | | Content Expansion | ✅ Complete | Examples, code samples, references | | UX & Community | ✅ Complete | Templates, checklists, interactive elements | | Maintenance Setup | ✅ Complete | Automation, governance, procedures | | Launch Preparation | 🔄 In Progress | Final polish, validation, promotion |

Success Metrics

• 📈 Adoption: 2+ major registries implement Universal Format
• 🌍 Community: 50+ contributors, active discussions
• 🎯 Impact: Referenced in OWASP Top 10, industry standards
• 📊 Quality: 100% link validity, comprehensive coverage

---

Resources

Documentation

• Full Website: Complete documentation
• Risk Details: All 10 risks in one document
• Security Checklist: Assessment and remediation guide
• Universal Format: Cross-platform standard specification

Research & References

• Snyk ToxicSkills (Feb 2026): Comprehensive skill ecosystem audit
• Check Point Research: Claude Code vulnerability analysis
• Antiy CERT: ClawHavoc campaign analysis
• CSA MAESTRO: 7-layer threat model for agentic AI

Tools & Automation

• Validation Script: Content quality checker
• Maintenance Guide: Long-term project procedures
• GitHub Actions: Automated testing and deployment

---

License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

You are free to share and adapt this material for any purpose, provided you give appropriate credit, provide a link to the license, indicate if changes were made, and distribute your contributions under the same license.

---

Contact

Project Lead: Ken Huang — OWASP AIVSS Lead, Agentic AI Security Researcher

• 📧 Email: [email protected]
• 🐙 GitHub: @kenhuangus
• 🔗 LinkedIn: Ken Huang

For questions, suggestions, or to get involved:

• Open an issue on GitHub

---

OWASP Agentic Skills Top 10 - Protecting the AI Agent Ecosystem 🛡️🤖

• Relationship to Existing OWASP Projects
• Getting Started
• Target Audience
• Project Status and Timeline
• Key Deliverables
• Scope
• Leadership and Governance
• Key Research and References
• License

---

Overview

The OWASP Agentic Skills Top 10 (AST10) documents the 10 most critical security risks in agentic AI skills across all major AI agent platforms. Skills represent the execution layer that gives agents real-world impact: they define not just what resources agents can access, but how they orchestrate multi-step workflows autonomously.

While significant attention has been devoted to securing large language models (LLMs) and the Model Context Protocol (MCP) tool layer, the intermediate behavior layer—embodied in agentic skills—has emerged as a particularly vulnerable and under-protected component of the AI agent ecosystem. This project exists to close that gap.

Mental Model: MCP = how the model talks to tools; AST10 = what those tools actually do.

MAESTRO Mapping: Each AST10 risk is mapped to the Cloud Security Alliance's MAESTRO 7-layer threat model for agentic AI systems, enabling targeted threat localization and cross-layer risk analysis.

---

The Problem: A Crisis Already in Progress

This is not a theoretical future risk. The AI agent skill ecosystem is under active attack as of Q1 2026.

By the numbers:

| Metric | Figure | Source | |--------|--------|--------| | Skills scanned | 3,984 | Snyk ToxicSkills (Feb 2026) | | Skills with security flaws | 1,467 (36.82%) | Snyk ToxicSkills (Feb 2026) | | Skills with critical issues | 534 (13.4%) | Snyk ToxicSkills (Feb 2026) | | Confirmed malicious payloads | 76+ | Snyk ToxicSkills (Feb 2026) | | ClawHavoc campaign: malicious skills | 1,184 | Antiy CERT (Feb 2026) | | OpenClaw instances internet-exposed | 135,000+ | SecurityScorecard (Feb 2026) | | CVEs disclosed (OpenClaw alone) | 9 (3 with public exploits) | Endor Labs (Feb 2026) | | Skills analyzed across all registries | 30,000+ | National CIO Review / Cisco (2026) | | Skills containing at least one vulnerability | >25% | National CIO Review (2026) |

The ClawHub registry—the primary marketplace for OpenClaw skills—became the first AI agent registry to be systematically poisoned at scale. Five of the top seven most-downloaded skills at peak infection were confirmed malware. The registry has since implemented automated scanning and partnered with VirusTotal, but the broader ecosystem remains largely unprotected.

Check Point Research disclosed two critical vulnerabilities in Claude Code (CVE-2025-59536, CVSS 8.7; CVE-2026-21852, CVSS 5.3) demonstrating that repository-level configuration files now function as part of the execution layer—simply cloning and opening an untrusted project can trigger remote code execution and API key exfiltration before any user consent dialog appears.

No comprehensive security framework or dedicated guidance for agent skills existed before this project. That gap is what AST10 addresses.

---

What Are Agentic Skills?

Agentic AI skills are reusable, named behaviors that encode complete workflows, including:

• Task understanding and goal decomposition
• Multi-step planning and tool orchestration
• File system, network, and shell access
• Safety guardrails and output formatting
• Persistent memory and cross-session state

Unlike MCP tools (which define what resources and actions are available), skills define how to use those tools in sequence to accomplish user goals. This behavioral abstraction layer creates unique security challenges that cannot be addressed by securing either the model or the protocol layer alone.

The "Lethal Trifecta" (Simon Willison / Palo Alto Networks, 2026): An AI agent skill is especially dangerous when it simultaneously has: 1. Access to private data (SSH keys, API credentials, wallet files, browser data) 2. Exposure to untrusted content (skill instructions, memory files, email) 3. Ability to communicate externally (network egress, webhook calls, curl)

Most production agent deployments today satisfy all three conditions.

Skill Formats by Platform

| Platform | Skill Format | Primary Risk File | |----------|-------------|-------------------| | OpenClaw | SKILL.md (YAML frontmatter + Markdown) | SKILL.md, SOUL.md, MEMORY.md | | Claude Code | skill.json / YAML + scripts/ | .claude/settings.json, hooks config | | Cursor / Codex | manifest.json + handler scripts | manifest.json, tool configs | | VS Code | package.json + extensions | package.json, extension.ts |

---

Incident Timeline (2026)

The following is a condensed timeline of confirmed real-world incidents involving AI agent skill security, drawn from publicly disclosed research and CVE records.

January 2026

• Jan 27–29: ClawHavoc campaign launches. Attackers register as ClawHub developers and flood the registry with 341 malicious skills in a 3-day window. All 335 AMOS-delivering skills share a single C2 IP (91.92.242[.]30). Target data includes exchange API keys, wallet private keys, SSH credentials, browser passwords, and .env files. Skills also write malicious instructions directly into MEMORY.md and SOUL.md for session-persistent backdooring.
• Jan 31: ClawHavoc surge peaks. Koi Security names the campaign and begins coordinated removal effort. Some packages persist for weeks.

February 2026

• Feb 1: Koi Security publishes first public ClawHavoc analysis.
• Feb 3: Snyk publishes "From SKILL.md to Shell Access in Three Lines of Markdown" threat model, documenting how three lines of markdown in a SKILL.md file can instruct an agent to read SSH keys and exfiltrate them.
• Feb 4: Alice publishes findings on several published OpenClaw skills found to be actively malicious while in use by over 6,000 users — detected via behavioral analysis (Yahoo Finance).
• Feb 5: Snyk publishes ToxicSkills — the first comprehensive security audit of the AI agent skill ecosystem. Key findings: 36% of skills contain security flaws; 13.4% contain critical-level issues; 76 confirmed active malicious payloads; 8 malicious skills still live at time of publication.
• Feb 5: Snyk publishes "280+ Leaky Skills: How OpenClaw & ClawHub Are Exposing API Keys and PII" — a parallel finding showing credential exposure at scale through over-permissioned skills.
• Feb 10: Snyk documents "How a Malicious Google Skill on ClawHub Tricks Users Into Installing Malware" — typosquatting and fake brand impersonation confirmed as active tactics.
• Feb 11: Snyk publishes "Why Your Skill Scanner Is Just False Security (and Maybe Malware)" — demonstrating that pattern-matching scanners miss the majority of critical threats, which rely on natural-language instruction manipulation rather than code signatures.
• Feb 14: OpenClaw patches log poisoning vulnerability (version 2026.2.13). Attackers could write malicious content to agent log files via WebSocket requests; since the agent reads its own logs for troubleshooting, injected text could influence decisions and trigger indirect prompt injection.
• Feb 25: Check Point Research publicly discloses CVE-2025-59536 (CVSS 8.7) and CVE-2026-21852 (CVSS 5.3) in Claude Code. Both were patched months earlier but the disclosure confirms: repository-controlled configuration files can silently execute arbitrary shell commands and exfiltrate API keys at project open time, before any trust dialog.
• Feb 26: ClawJacked disclosed by Oasis Security (CVE-2026-28363, CVSS 9.9). Malicious websites can brute-force localhost WebSocket connections with no rate limiting to silently hijack local OpenClaw instances, register new devices without user prompts, and exfiltrate data through existing agent integrations. OpenClaw patches within 24 hours (version 2026.2.25).
• Feb 2026: Antiy CERT publishes ClawHavoc Campaign Analysis, classifying malware as Trojan/OpenClaw.PolySkill. Final tally: 1,184 malicious skills across 12 publisher accounts. Hudson Rock separately identifies Vidar infostealer variants specifically targeting OpenClaw agent identity files (openclaw.json, device.json, soul.md, memory.md).
• Feb 2026: Microsoft Defender Security Research Team issues advisory: "Because of these characteristics, OpenClaw should be treated as untrusted code execution with persistent credentials. It is not appropriate to run on a standard personal or enterprise workstation."
• Feb 2026: BlueRock Security analyzes 7,000+ MCP servers; finds 36.7% potentially vulnerable to SSRF. Proof-of-concept against Microsoft's MarkItDown MCP server retrieves AWS IAM keys from EC2 metadata endpoint.

March 2026

• Mar 2026: SecurityScorecard confirms 135,000+ OpenClaw instances publicly internet-exposed with insecure defaults; 53,000+ correlated with prior breach activity. Bitdefender telemetry confirms employees deploying OpenClaw on corporate devices with no SOC visibility.
• Mar 2026: Snyk and Tessl announce registry-level skill security scanning partnership. Snyk and Vercel previously partnered to scan skills on skills.sh at install time.
• NIST / CAISI: Federal Register RFI on AI Agent Security (published Jan 8, 2026, comments closed Mar 9, 2026) — the first formal US government solicitation specifically addressing AI agent security risks.

---

The 10 Risks — Summary Table

| # | Risk | Severity | Platforms Affected | Key Mitigation | Real-World Evidence | |---|------|----------|--------------------|----------------|---------------------| | AST01 | Malicious Skills | Critical | All | Merkle root signing, registry scanning | ClawHavoc (1,184 skills), ToxicSkills (76 payloads) | | AST02 | Supply Chain Compromise | Critical | All | Registry transparency, provenance tracking | ClawHub collapse, Claude Code CVE-2025-59536 | | AST03 | Over-Privileged Skills | High | All | Least-privilege manifests, schema validation | 280+ credential-leaking skills (Snyk, Feb 2026) | | AST04 | Insecure Metadata | High | All | Static analysis, safe parsers, sandboxed loading | Fake "Google" skill impersonation; YAML payload delivery in SKILL.md | | AST05 | Untrusted External Instructions | High | All | Source inventory, content pinning and inlining, continuous rescanning | Anthropic docs warn fetched URLs "may contain malicious instructions"; Air's "Story of Skills" PoC bypassed all scanners and proved a possible takeover of 26,000 agents via untrusted external instructions | | AST06 | Weak Isolation | High | All | Containerization, Docker sandboxing | OpenClaw host-mode execution, 135K exposed instances | | AST07 | Update Drift | Medium | All | Immutable pinning, hash verification | ClawJacked (CVE-2026-28363), patch-lag exploitation | | AST08 | Poor Scanning | Medium | All | Semantic + behavioral multi-tool pipeline | Pattern-matcher bypass via natural-language injection | | AST09 | No Governance | Medium | All | Skill inventories, agentic identity controls | 53K exposed instances with no SOC visibility | | AST10 | Cross-Platform Reuse | Medium | All | Universal YAML format | Malicious skills ported across ClawHub, skills.sh |

---

Detailed Risk Descriptions

Each of the 10 risks is documented in a separate file. Click on the risk name to view the full details.

| # | Risk | Severity | Platforms Affected | Key Mitigation | Real-World Evidence | |---|------|----------|--------------------|----------------|---------------------| | AST01 | Malicious Skills | Critical | All | Merkle root signing, registry scanning | ClawHavoc (1,184 skills), ToxicSkills (76 payloads) | | AST02 | Supply Chain Compromise | Critical | All | Registry transparency, provenance tracking | ClawHub collapse, Claude Code CVE-2025-59536 | | AST03 | Over-Privileged Skills | High | All | Least-privilege manifests, schema validation | 280+ credential-leaking skills (Snyk, Feb 2026) | | AST04 | Insecure Metadata | High | All | Static analysis, safe parsers, sandboxed loading | Fake "Google" skill impersonation; YAML payload delivery in SKILL.md | | AST05 | Untrusted External Instructions | High | All | Source inventory, content pinning and inlining, continuous rescanning | Anthropic docs warn fetched URLs "may contain malicious instructions"; Air's "Story of Skills" PoC bypassed all scanners and proved a possible takeover of 26,000 agents via untrusted exte