AI Agents Are Now Running Hands-On Cyber Attacks, and They Leave a Noisy Trail

Wiz Research says its incident response and threat hunting work shows a clear trend: intrusions where AI agents appear to carry out much of the hands-on work, and possibly orchestrate the attack itself. These agents run through software harnesses that write custom tools, check the results, adapt and try again at machine speed. A human attacker still sets the objectives and guardrails.

The researchers note these attacks span the range from sharply targeted to broadly opportunistic. A single exposed service can lead to stolen cloud credentials, stolen Kubernetes tokens, movement across systems and bulk theft of secrets. Escalation that once demanded skilled manual effort is now compressed by AI. Agents often try several approaches in parallel, which makes the attacks verbose.

There is some good news for defenders. Current agentic attacks tend to be noisy and leave a broad trail of observable activity. The difficulty is telling this activity apart from legitimate automation and linking individual signals into a clear picture of an attack. Wiz points to recent public disclosures involving named frameworks such as ARTEX and Hermes, plus custom-built workflows and cases where the tooling is unknown. This is only the opening of the research, which also looks at attribution, attack scale and the future of defence.