Multi-agent security

Security for collaborating AI agents

Classical controls harden individual models or infrastructure. Multi-agent attacks exploit trust between agents: delegation chains, tool handoffs, and data passed across services without a single choke point.

What Burgus observes

• Per-request audit metadata on LLM proxy traffic (vendor, model, tokens, timing)
• Workflow and session rollups when X-Llmproxy-* correlation headers are used
• Agent interaction graphs — who talked to whom inside a workflow
• Platform baseline and custom policy alerts mapped to OWASP LLM Top 10 and MITRE ATLAS
• Behavioral baselines for volume, graph fan-out, loops, and coordinated bursts

MCP reporting

audit_multi_agent_security_report combines traceability coverage, content-policy alerts, and behavioral anomaly rollups in one tenant-scoped view — designed for security engineers triaging multi-agent incidents.