AI agents carried out a ransomware attack in under 10 hours

A ransomware attacker used frontier AI models and a custom agentic framework to breach an enterprise network in under 10 hours, work that would typically take a human team about two weeks, according to Unit 42, Palo Alto Networks' incident response arm.

The attacker's agents chained more than 50 MITRE ATT&CK techniques without step-by-step human direction: automated reconnaissance to map the target's internal microservices, credential theft from hardcoded tokens in code repositories and a secrets-management system, and lateral movement into cloud, identity, and CI/CD environments. The agents also tried planting a backdoor in the victim's Terraform configuration, a move branch protection caught before it landed.

Unit 42 said the agents "monitored, evaluated, acted and re-planned in real time" throughout the intrusion, running parallel calls to multiple AI models and passing state between them as structured Markdown files. Once inside, the attacker turned the victim's own cloud AI infrastructure into part of the attack chain, using stolen API keys to invoke hosted models for further automation.

Before negotiating a ransom, the attacker had the agents compile an 80-page technical writeup of the exploited weaknesses and hand it to the victim, an unsolicited penetration-test report attached to the extortion demand.

For teams running public API endpoints, CI/CD pipelines, and secrets managers without a dedicated security function, the hardening checklist hasn't changed: rotate and scope credentials tightly, keep branch protection on infrastructure-as-code repos, and don't assume a public endpoint is too minor to map. Agentic tooling has cut the time an attacker needs to find and chain those gaps down to hours.