Tool Security Advisory (TSA)

The Model Context Protocol (MCP) enables AI agents to invoke tools that can read files, access networks, execute processes, and integrate with external services. These tools introduce security risks that are not fully captured by traditional vulnerability formats. This paper specifies the Tool Security Advisory (TSA) standard, a strict JSON advisory format with a feed model and trust framework designed for MCP tool ecosystems. TSA adds MCP-specific primitives such as semantic drift, capability abuse, agent execution context, and Tool Bill of Materials (TBOM) content hash binding. TSA also defines deterministic canonicalization and cryptographic signing for integrity and provenance, and provides interoperability with OSVthrough loss-minimized mappings. A reference implementation and SDK are available for evaluation and early adoption.