AVE: A Behavioral Vulnerability Standard for Agentic AI
AVE is a new standard classifying behavioral vulnerabilities in agentic AI, using AIVSS scoring and the bawbel-scanner reference tool for CI checks.
AVE introduces a classification standard for the natural-language artifacts that agentic AI systems actually execute — skill files, MCP server manifests, and system prompts. Existing standards like CVE and OSV map flaws to packages and version ranges, but they can't describe a prompt injection buried in an MCP tool description: there's no package, no version, just behavior that an LLM reads and acts on.
AVE closes that gap by assigning stable identifiers to distinct behavioral classes in agentic AI, scoring each with OWASP AIVSS v0.8, and mapping records to OWASP MCP Top 10 and MITRE ATLAS so findings fit into frameworks security teams already use. The current catalog holds 59 records: 1 critical, 12 high, 44 medium, and 2 low severity.
AVE itself is a standard, not a product — bawbel-scanner serves as its reference implementation, letting teams scan skill files and MCP server cards in CI or pre-commit hooks and block risky payloads before deployment. The practical takeaway for engineers: dependency scanning alone no longer covers the attack surface once an LLM is the runtime interpreting your instruction files.