Attackers Use LLM Agent for Post-Exploitation After Marimo CVE-2026-39987 Exploit
An unknown threat actor exploited CVE-2026-39987 in Marimo to gain initial access, subsequently using a large language model (LLM) agent to conduct post-exploitation activities, including stealing credentials and exfiltrating a PostgreSQL database. This attack highlights the evolving tactics of cybercriminals and the need for proactive security measures, particularly around vulnerable software and credential management.
The incident, documented by Sysdig on May 10, 2026, began with an attacker leveraging the publicly disclosed vulnerability in Marimo (CVE-2026-39987) to compromise an internet-reachable notebook. Following this initial breach, the attacker extracted cloud credentials, obtained an AWS Secrets Manager SSH private key, and used it to initiate SSH sessions against a downstream bastion server. The attacker then swiftly exfiltrated the contents of an internal PostgreSQL database within under two minutes. The vulnerability itself allows unauthenticated attackers to execute arbitrary system commands, and while patched in version 0.23.0, the exploit was actively being used.