The era of AI-driven cyberattacks is no longer a future threat it is happening right now. On May 10, 2026, the Sysdig Threat Research Team (TRT) witnessed a chilling evolution in cyber warfare.
Attackers deployed an autonomous Large Language Model (LLM) agent to execute a flawless post-exploitation routine, ultimately exfiltrating an internal PostgreSQL database in under two minutes.
The intrusion began with a known vulnerability, CVE-2026-39987, affecting internet-exposed marimo notebooks. Attackers exploited this Remote Code Execution (RCE) flaw to harvest cloud credentials from the compromised host.
Instead of manually sifting through the data, the attackers fed these credentials into an AI agent.
The AI rapidly replayed the stolen keys through a fanned-out network of Cloudflare Workers, retrieved an SSH private key from AWS Secrets Manager, and launched parallel sessions against a downstream SSH bastion server.
LLM Agent Pivots Database
According to Sysdig TRT’s Senior Director, Michael Clark, “We are not watching AI replace attackers. We are watching attackers replace their scripts with AI.” This incident perfectly illustrates the shift from rigid, pre-built attack scripts to highly adaptable AI orchestrators.
Sysdig identified four distinct markers proving this attack was driven by an LLM in real-time rather than a human or a traditional script.
Improvised Target Analysis The agent dumped specific database tables without prior knowledge of the schema, relying on AI guesswork to infer common application structures.
PGPASSWORD=<harvested-from-pgpass> psql -h internal-db -U app -d app -c \
'SELECT * FROM credential;' 2>&1 | head -40
Leaked Planning Monologue The command stream accidentally captured a Chinese comment, translated as “See what else we can do,” fired at sub-second speed across multiple IPs.
Machine-Optimized Formatting The agent used command separators, bounded line captures, and discarded error messages to keep its own context window clean and easy to process.
Self-Consuming Output The AI instantly read the output of its previous commands, lifted the valuable credentials, and fed them directly into the next stage of the attack.
This adaptiveness means attackers no longer need to write custom playbooks for every target. The AI agent observes the environment, decides the best path forward, and executes it flawlessly.
According to Sysdig research, the speed of this attack highlights why immediate patching is critical. Once the AI agent took over, the pivot from initial access to full database exfiltration happened in a matter of minutes.
To protect your infrastructure from this emerging AI-driven threat, security teams must implement the following mitigation strategies immediately.
Update marimo instances to version 0.23.0 or later to patch the RCE vulnerability. Restrict network access to the /terminal/ws endpoint if patching is immediately impossible.
Rotate all AWS credentials, database passwords, and SSH keys that were stored on publicly accessible marimo instances. Audit environment variables and .env files for unauthorized access or exposed secrets.
Deploy comprehensive runtime threat detection to catch lateral movement, as traditional signature-based detections often fail against adaptable AI agents.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.