nullhypeai/agentic-threat-exposure-checklist

This repository is a free, one-page security checklist for anyone running AI agent tools or infrastructure. It was written after studying a real incident called JADEPUFFER, described as the first ransomware attack carried out end to end by an autonomous LLM agent with no human operator, documented by Sysdig Threat Research in July 2026. Each item in the checklist maps directly to a step that attack actually used, so it is meant to be a practical, evidence-based list rather than generic advice. The checklist is organized into five areas. The first is locking down internet-facing AI tools, since the attacker got in through a publicly reachable Langflow instance with a known, patchable vulnerability, the advice here is to inventory every AI framework or agent dashboard you run and make sure none of them sit directly on the public internet. The second is protecting LLM API keys, since a stolen key gives an attacker free compute, so the checklist calls for scoping keys tightly, killing unused ones, and treating unusual token spend as a security alert rather than just a billing issue. The third covers stopping attackers from moving sideways once inside, mainly by removing default and reused credentials on databases and other services, since the attacker in this case used well-known default logins to reach a production database. The fourth is about recovery, noting that in this attack the encryption key was never stored anywhere, so paying a ransom would not have helped, and it recommends actually testing a full restore rather than just confirming backups exist. The fifth is about detection speed, pointing out that the attacker adapted to a failed login in 31 seconds, so it argues for behavior-based detection instead of relying on recognizing reused malware files. The document ends with five direct questions meant to be asked of a security team on a Monday morning, covering AI tool inventory, spend alerts, tested recovery time, ransom policy, and detection speed. It is free to copy, fork, and share with a team, with no signup required, and it credits Sysdig Threat Research and public vulnerability databases as its sources.