An AI-Assisted Cyber Attack: Inside a Unit 42 Investigation

Unit 42 investigated an intrusion in which a human adversary used frontier AI agents and agentic attack frameworks to compress weeks of intrusion tradecraft into ~10 hours, breaching a public-facing service, harvesting secrets from code repositories and a secrets manager, hijacking CI/CD pipelines, and abusing stolen cloud keys to commandeer the victim's AI endpoints. The report maps observed techniques to MITRE ATT&CK and ATLAS, lists agentic indicators (e.g., structured Markdown, parallel LLM calls), and recommends defenses including automated containment playbooks, strict governance of AI endpoints and API keys, detection of behavioral loops, and hardening DevOps pipelines.