TLDR Tech / DevOps
AI agent at the wheel: How an attacker used LLMs to move from a CVE to an internal database in 4 pivots
The Sysdig Threat Research Team observed what appears to be the first documented AI agent-driven cyberattack on May 10, where an attacker exploited a marimo notebook vulnerability (CVE-2026-39987) and used a large language model to autonomously navigate from initial access through AWS credentials to exfiltrating an entire PostgreSQL database in under two minutes. Four key signatures pointed to real-time AI composition rather than pre-scripted automation: the agent dumped a non-existent "credential" table based on schema assumptions, left a Chinese-language internal monologue comment mid-attack, used distinctively AI-formatted commands with separators and bounded captures, and dynamically chained outputs from one command as inputs to the next—all while spreading requests across multiple Cloudflare Workers IPs to evade detection.
Read the full story