Frontier AI agents just breached an enterprise in 10 hours. They left behind an 80-page security audit.

Unit 42 documented the first multi-agent frontier AI attack on a real enterprise network.

The attacker set an objective. Then stepped back.

A coordinated fleet of AI agents ran the entire operation — reconnaissance, credential harvesting, privilege escalation, CI/CD exploitation, cloud infrastructure takeover — across cloud, identity, DevOps, and SaaS planes simultaneously.

Work that would take a human red team roughly two weeks was compressed into under 10 hours. More than 50 MITRE ATT&CK techniques. No zero-days used.

The agents harvested hard-coded tokens directly from code repositories. Hijacked CI/CD workflows. Exfiltrated cloud access keys. Then turned the victim's own AI infrastructure into post-compromise compute — hiding orchestration traffic inside legitimate model calls.

The final artifact: an 80-page technical security audit documenting every vulnerability exploited. Delivered by the same agents that just dismantled the company's defenses.

The attacker disclosed during ransom negotiations that they used frontier AI models and purpose-built agentic frameworks.

One control stopped them: mandatory multi-party code review on infrastructure-as-code repos. The agents tried to plant Terraform backdoors. Branch protection required human approval. That single gate held.

If your incident response runbook handles cloud, identity, and CI/CD compromise sequentially — you are not calibrated for machine-speed attacks. Containment must be simultaneous across all planes. Automated. No human triage bottleneck.

Audit your code repositories for hard-coded credentials today. Test whether your containment playbooks are synchronized. If they aren't, you are in the exact risk window this attack exploited.