Description

A threat actor used frontier artificial intelligence models and specialized agentic frameworks to compromise an enterprise environment, obtain privileged credentials, and misuse cloud-based AI resources in less than 10 hours. The investigation by Palo Alto Networks Unit 42 demonstrates how AI-assisted automation can dramatically accelerate intrusion operations. Activities that could traditionally require several skilled operators and weeks of coordinated effort were completed rapidly through multiple specialized AI agents operating under the attacker’s direction. The intrusion began with the exploitation of a publicly accessible web service, providing the attacker with an internal network pathway for reconnaissance and follow-on operations. AI agents automatically mapped internal services and microservices before searching source-code repositories for exposed passwords, API tokens, service credentials, and cloud access keys. The discovered secrets enabled access to the organization’s secrets-management platform, where administrative and root-level credentials were obtained. The agents subsequently supported privilege escalation, lateral movement, persistence, and data collection while adapting their actions when defensive controls disrupted specific techniques. Unit 42 observed more than 50 MITRE ATT&CK techniques, along with parallel LLM queries, structured Markdown files for maintaining operational context, and AI-assisted automation scripts. The attacker also targeted the organization’s DevOps infrastructure, attempting to extract cloud credentials and modify Terraform configurations to establish persistence. Strong branch-protection controls prevented the infrastructure-as-code changes, highlighting the importance of protected repositories and mandatory reviews. Stolen cloud credentials were later used to access the victim’s AI endpoints and consume its computing resources. Organizations should treat AI infrastructure as critical enterprise assets and implement least-privilege access, secret scanning, rate limiting, detailed logging, and continuous monitoring across AI, identity, cloud, and DevOps environments. Incident-response procedures should support rapid credential revocation, session termination, pipeline suspension, and cloud-resource isolation.