JADEPUFFER: ransomware intrusion assessed to have been executed end-to-end by an LLM agent via Langflow RCE

One victim environment; 1,342 Nacos configuration items encrypted and multiple databases dropped, unrecoverable

Sysdig's 'first ransomware run start to finish by an AI agent' framing is inference, not proof: they never observed the model or agent framework. Their evidence is behavioural - self-narrating payloads, a 31-second failure-diagnosis-and-correction cycle, semantic parsing of free-text environment data, and a ransom Bitcoin address (3J98t1WpEZ73CNmQviecrnyiWrnqRhWNLy) that is the canonical documentation example address, which may be an LLM hallucination. Sysdig themselves flag that exfiltration is claimed only in the agent's own code comments and was not independently observed. Note the irony that the entry point was an AI agent-builder platform. | Merged from duplicate id `2026-07-sysdig-jadepuffer`. Also reported as: 'JADEPUFFER: an end-to-end LLM-driven ransomware operation that extorted a production database'.