Assisted, Generated, Operated: Inside Three Real AI Attacks

This session is based on three real incidents that Bitdefender investigated - actual intrusions, not malware samples reverse-engineered in isolation. Different things get called “AI attacks,” and these real incidents show they are not the same threat. Together they support a classification of external AI-enabled threats, separated by what the model actually does in an operation: supports the operator, produces the artifact, or runs the attack.

SilkParasite, a China-nexus actor, illustrates AI-assisted operations: models supporting reconnaissance, social engineering and tooling. APT36, a Pakistani actor, illustrates AI-generated vibeware - tooling that was model-written, and that proved derivative, bloated and readily detected in the field. The third case, a 3AM ransomware intrusion, illustrates AI-operated activity, where the operation itself was agent-driven.

For each tier we show how to recognize it, what it leaves behind in your telemetry and detections, and how worried you should actually be. Drawing on Bitdefender Labs and MDR data, I’ll give a verdict on each: a gimmick, an experiment, or something you should expect more threat actors to adopt in 2027.