AI agents carried out every step of this ransomware attack – then left the victim an 80-page security audit

In a swift and highly efficient ransomware attack, a human attacker utilized frontier AI models to infiltrate an enterprise network in under 10 hours, a task typically taking human operators approximately two weeks, according to Unit 42. The attacker claimed to have employed frontier models and agentic attack frameworks, with AI agents executing each step of the intrusion, including leaving an extensive 80-page security audit for the victim company.

The efficiency of the attack was attributed to AI-assisted operational efficiency. The attacker handed tactical execution to AI agents that continuously monitored, evaluated, acted, and re-planned in real time, accelerating the attack chain. The security firm did not reveal the specific AI models and frameworks utilized by the attacker.

The attack commenced with the human operator using AI agents for reconnaissance, followed by gaining access through a public API endpoint. Once inside, the attacker deployed an automated recon agent to map internal microservices. Additionally, subagents scraped code repositories to obtain hard-coded tokens and service passwords. Using these stolen credentials, AI intruders accessed the organization's secret-management system, eventually obtaining root system access.

Subsequent "specialist pivot agents" verified access to the company's cloud, identity, CI/CD, container, and SaaS environments. The attacker also manipulated CI/CD workflows to steal cloud access keys, converting the victim's cloud AI services into post-compromise infrastructure. This action enabled the attacker to utilize the victim's compute resources while camouflaging orchestration traffic within legitimate activity.

Upon achieving the attacker's objectives, an AI agent presented the victim company with an 80-page report detailing numerous security vulnerabilities and exploited findings. Palo Alto Networks emphasized that the only way to protect against machine-speed attacks is to employ AI agents themselves. The incident response team recommended that companies treat AI as core infrastructure, involving inventorying every model endpoint, API key, Model Context Protocol (MCP) gateway, and AI tool integration, implementing rate limits and least-privilege policies to avoid unexpected and large token bills.