AI Agent Exposed Its Own Hacking Operation by Leaving Logs on Public Server

An autonomous AI-driven cyberattack targeting internet-exposed automation tools including Langflow, n8n, and Marimo was uncovered after the agent accidentally hosted its own tool logs and activity on a publicly accessible web server. The attack was launched via a single Telegram command and used an LLM-based agent framework to independently select targets, choose exploits, and adapt when attempts failed. Despite the novel decision-making layer, most exploitation attempts were unsuccessful due to configuration mismatches, and the operation was compromised by basic operational security failures. Security analysts note the incident is less a sign of sophisticated AI hacking and more a reflection of how attackers are delegating repetitive exploitation tasks to error-prone AI models. The broader takeaway is that dev and automation tools routinely left exposed to the internet remain a persistent vulnerability, now increasingly targeted at scale with lower effort than before.