Skill CVE Database¶
ZIRAN maintains a curated database of known security vulnerabilities in popular AI agent tools and skills.
What is a Skill CVE?¶
The database contains two types of entries:
- Real CVEs — Verified vulnerabilities from the NVD/GitHub Advisory Database (e.g.
CVE-2025-68664) - Design Risks — Architectural weaknesses documented in OWASP LLM Top 10 and framework security guides (e.g.
DESIGN-RISK-001)
ID Format¶
CVE-YYYY-NNNNN # Real CVE from NVD
DESIGN-RISK-NNN # OWASP/framework design risk
Current Database¶
ZIRAN ships with 27 seed entries covering:
Real CVEs¶
- LangChain — CVE-2023-46229 (SSRF), CVE-2025-68664 (serialization injection), CVE-2025-65106 (template injection), CVE-2025-46059 (prompt injection), CVE-2025-8709 (SQL injection), CVE-2025-64439 (RCE), CVE-2025-2828 (SSRF)
- MCP — CVE-2025-53109 (symlink bypass), CVE-2025-53110 (path traversal), CVE-2025-68145 (git path bypass), CVE-2025-6514 (command injection), CVE-2026-27825 (file write RCE)
- Other — CVE-2025-32711 (M365 Copilot), CVE-2026-29783 (Copilot CLI)
Design Risks (OWASP LLM Top 10)¶
- LangChain — ShellTool, RequestsGetTool, ReadFileTool, WriteFileTool, QuerySQLDataBaseTool, PythonREPLTool, SerpAPIWrapper, ConversationBufferMemory, GmailSendMessage, TavilySearchResults
- CrewAI — ScrapeWebsiteTool, FileReadTool, CodeInterpreterTool
Checking Your Agent¶
from ziran.application.skill_cve import SkillCVEDatabase
db = SkillCVEDatabase()
matches = db.check_agent(discovered_capabilities)
for cve in matches:
print(f"{cve.cve_id}: {cve.skill_name}")
print(f" Severity: {cve.severity}")
print(f" Type: {cve.risk_type}")
print(f" Risk: {cve.description}")
print(f" Fix: {cve.remediation}")
Submitting a Skill CVE¶
Found a vulnerability in an agent tool? Help the community by submitting it:
- Open a GitHub issue using the Skill CVE template
- Include:
- Tool/skill name and version
- Framework (LangChain, CrewAI, etc.)
- Vulnerability type and severity
- Description and proof of concept
- Remediation guidance
- Real CVE ID if available (from NVD or GitHub advisories)
- The ZIRAN team will review and add it to the database
Vision¶
We envision a community-maintained database of agent tool vulnerabilities — similar to how the traditional CVE system works, but focused on the unique risks of AI agent tools.
As the ecosystem grows, this database will become an essential resource for: - Developers — Know the risks of the tools you're giving your agents - Security teams — Audit agent deployments against known vulnerabilities - Researchers — Track and document new vulnerability patterns