The holder of Operational Threat Hunting for AI Systems possess the capability of the following demanding skills :
- Map the enterprise AI stack and identify the log sources and attack surface at each layer.
- Define logging and observability requirements for agents, RAG pipelines, vector stores and MCP tools.
- Detect rogue agent registration and agent card spoofing from A2A registration, discovery and routing events.
- Correlate agent-to-agent activity using OpenTelemetry trace and correlation IDs.
- Identify data poisoning and indirect prompt injection in RAG ingestion and retrieval logs.
- Spot suspicious vector database activity, such as enumeration queries and embedding model or dimensionality changes.
- Uncover MCP tool poisoning through tool registration, invocation and cross-tool data flow analysis.
- Detect covert exfiltration hidden in tool outputs, such as Base64 or steganographic payloads.
- Write KQL queries in Kibana/Elasticsearch to filter, pivot and correlate AI telemetry.
- Map AI attacks to MITRE ATLAS, MITRE ATT&CK and OWASP Top 10 for LLM Applications.
- Reconstruct AI incident timelines and report findings with detection recommendations.