From Packets to Intent: Hunting Adversaries in AI Telemetry

As AI systems become part of critical products and workflows, they introduce a new security surface where attacks happen through language. In traditional security domains, threat hunting focuses on signals such as network ports, traffic patterns, or system activity. In AI security, the signals are different. Instead of packets and processes, defenders analyze text interactions with models to identify malicious intent.

Effective threat hunting in AI systems requires more advanced tools. Signals hidden within natural language often require analyzing text using tools such as embedding models and perplexity to surface suspicious intent and anomalous behavior. In this talk we demonstrate a novel approach for conducting effective threat hunting in AI driven applications.

AI security changes the defender’s job, the attack surface is no longer limited to hosts, identities, and network traffic. When language becomes the interface to business logic, data access, and automated actions, malicious behavior can look like normal user interaction unless you know what to look for.

This talk focuses on threat hunting in AI systems from a practical security perspective. It examines the signals defenders can use when investigating text driven attacks, including prompt structure, semantic similarity, anomalous intent, embeddings, perplexity, and suspicious workflow patterns across models, tools, and retrieval layers.

The talk will also cover concrete attack scenarios such as prompt injection, abuse of agent capabilities, and attempts to extract sensitive information through model interaction. The goal is to show how defenders can move from generic AI security concerns to usable hunting methods and detection strategies that work in production environments.