agent-bom runtime

This Skill addresses the pain points of runtime security monitoring for AI agents. In practical operations, the contextual interactions and audit logs of AI Agents are scattered, making it difficult to quickly identify security threats such as lateral movement and privilege abuse. Traditional logging tools lack correlation analysis of Agent behavior graphs and cannot automatically link runtime logs with CVE vulnerability databases. This tool helps security teams detect risks in real-time during AI system operation by constructing context graphs, correlating audit logs with CVE discoveries, and providing vulnerability analysis queries.

Usage is straightforward: simply install the agent-bom package via pipx, then invoke three core tools through natural language commands: context_graph (constructs Agent context graphs and analyzes lateral movement), runtime_correlate (cross-references runtime audit logs with CVE discoveries), and analytics_query (queries vulnerability trends, historical states, and runtime events). For example, instruct the AI with "Analyze the current agent's runtime audit logs, correlate with known CVEs, and generate a context graph." It will automatically parse JSONL-format audit files and output visualized context graphs and risk reports. Optionally integrate ClickHouse for persistent analysis—no API keys required.

It is well-suited for teams or individuals needing to protect AI agent security, especially operations engineers, security analysts, and agent developers. If you have deployed multiple AI Agents and are concerned about lateral movement, privilege abuse, or unknown vulnerabilities between them, this tool helps transform scattered audit logs into queryable security intelligence. It is also suitable for security auditing of runtime behavior during the development and testing phases of AI applications, ensuring risks are mitigated before deployment.

It is recommended to run a full context graph analysis and CVE correlation scan after each AI Agent version update or major configuration change. Note that it primarily relies on Python 3.11+ environments, and ClickHouse is an optional component—if not configured, analysis results are stored only in memory and cannot be retained long-term. Additionally, audit log files must be in JSONL format (from agent-bom proxy); other formats require conversion first.