Agentic AI Malware: Why the Cybersecurity Battle Isn’t Over

The talk starts from a puzzle: headlines describe AI malware that adapts to any defense, yet sample volumes and telemetry show few confirmed cases. The speaker separates malware that was merely written with an LLM from malware that calls a model while it runs. Examples include scripts with leftover AI-style comments, jailbroken or uncensored models, and a recent campaign that used a hosted code model to generate reconnaissance commands during intrusions.

The middle section presents a proof-of-concept PowerShell loader that carries no fixed payload. Each run it asks a hosted model for a command, executes it, feeds errors back, and stores the working prompt in the registry. The speaker then adds an orchestrator, research agents and tool agents that communicate locally. In testing against several commercial EDR products, the evasion attempts were ultimately detected, and the demo showed how repeated, varied attempts leave noticeable patterns.

The conclusion is that AI mainly speeds up planning, orchestration and exploitation at scale, and makes evasion more dynamic, but it does not yet give attackers a decisive edge. Basic defenses still work: behavior-based detection, file reputation, and monitoring of traffic to LLM services. Practitioners should expect an AI-versus-AI contest rather than a collapse of existing controls.