Tool: AgentSonar

AgentSonar is a network‑visibility tool that identifies likely LLM/AI agent traffic by correlating process ownership of sockets with contacted domains and applying a heuristic classifier that outputs an AI score between 0 and 1.

Summary

AgentSonar records outbound connections, associates them with processes via socket ownership, extracts domain indicators from TLS SNI and DNS, and produces scored events for each process→domain pair. Known agents can be defined to produce deterministic matches; domains marked as noise are excluded from scoring.

How it works (conceptual)
• Socket correlation: associates OS socket ownership with userland processes to reveal which binary initiated a connection.
• Domain extraction: uses TLS SNI and DNS observations as the domain identifier for each outbound flow.
• Heuristic classifier: analyzes traffic shape characteristics — byte/packet asymmetry, prevalence of small packets, long‑lived or streaming connections, and programmatic TLS patterns — to infer whether a flow resembles LLM API traffic.
• Scoring model: emits an AI-likelihood score between 0 and 1 per process→domain pair; known agents map to score 1.0, noise maps to 0.

Capabilities and workflows

AgentSonar provides persistent event storage and a triage-oriented workflow for reviewing high‑scoring unknowns and labeling them as agents or noise. It supports importing pre-built event streams for classification and encourages community submissions of agent classifications to improve coverage.

Limitations and scope

The approach relies on observable network metadata (socket ownership, SNI, DNS) and traffic-shape heuristics; encrypted payloads and obfuscated patterns remain outside content-level analysis. Deterministic detection depends on maintained known-agent mappings; heuristic scoring produces probabilistic indicators rather than definitive attribution.

Practical context

AgentSonar targets defenders seeking endpoint-to-domain visibility with AI‑specific signal enrichment, enabling detection of shadow AI usage where traditional allowlists may miss programmatic LLM traffic.

agentsonar #llm_detection #network_visibility #knostic

Source: https://github.com/knostic/AgentSonar/