AI agent security needs a composition graph, not just an SBOM

12 hours ago

The primary risk in AI agents lies not in individual plugins or packages but in the composition of components.
SCA scanners identify vulnerable packages but cannot correlate them with the agent's capabilities like reading private data or sending messages.
Example: the 'imessage' plugin combines an MCP server, skills, and npm packages to read/send messages, creating a risk surface through composition.
Analysis of Claude's plugin marketplace revealed 124 vulnerabilities concentrated in four messaging plugins, highlighting the correlation between vulnerable code and sensitive capabilities.
SCA sees only packages; runtime monitoring sees behavior; neither assesses the declared agent composition (plugins, skills, servers, permissions, etc.).
Agent security requires a composition graph view, mapping how components interconnect, to evaluate risks like untrusted input combined with data access.
OpenACA is an open-source tool that inventories agent stacks, attributes advisories to components, flags posture issues, and exports an Agent BOM.
Future development aims for graph-derived exposure analysis to prioritize findings based on plausible impact paths.
Agent security must shift from package-centric to composition-aware analysis, treating the agent as the primary unit of risk assessment.

Hasty Briefsbeta