An AI agent just ran a full ransomware attack in 31 seconds. Your incident response plan is built for humans.

An AI agent called JADEPUFFER just executed the first documented end-to-end ransomware operation without a human at the keyboard.

No operator. No manual exploitation. No human fixing failed payloads.

Sysdig's Threat Research Team captured the full chain:
Exploited a known Langflow vulnerability (CVE-2025-3248) that was patched over a year ago.

Harvested credentials across AWS, Azure, GCP, Alibaba, Tencent, OpenAI, Anthropic in parallel.

Pivoted to a production database. Created a backdoor admin account.

Login failed. 31 seconds later, the AI diagnosed the problem, rewrote the code, and succeeded.

600+ coordinated payloads delivered autonomously.

The attack scripts contained natural-language comments explaining why each step was being taken. LLM-generated code does this by default. Human operators don't.

The encryption key was generated, printed once, never stored. The victim cannot recover even if they pay.

None of the individual techniques were novel. CVE-2025-3248 was patched in May 2025. The Nacos bypass dates to 2021. The default JWT signing key has been public since 2020.

What's new is that an AI agent chained these neglected vulnerabilities into a complete extortion operation. The skill floor for running ransomware just dropped to whatever it costs to run an agent.

Audit your internet-facing AI infrastructure today. Patch Langflow. Rotate Nacos keys. If your incident response runbook assumes a human attacker with predictable paths, rewrite it.