Autonomous AI Agent Hijacks LLMs for Cyber Attacks

• •An autonomous AI agent can hijack exposed LLMs for cyber attacks.
• •The agent can perform up to 100,000 malicious actions without human oversight.
• •Attackers disguise credential theft as legitimate work using a jailbreak method.

Security researchers from Symantec have uncovered an autonomous AI agent capable of hijacking exposed large language models (LLMs) to conduct cyber attacks without human intervention. This agent can execute up to 100,000 malicious actions, including scanning for targets, exploiting vulnerabilities, stealing credentials, and deploying cryptocurrency miners. The agent was identified through an AI honeypot, which recorded over 127,000 requests from various addresses. The captured instruction set revealed that attackers used a jailbreak to bypass safety controls of commercial AI models, disguising credential theft as legitimate development work. This activity, termed 'LLMjacking,' allows attackers to utilize the computing power of hijacked models while shifting costs to the victims. The findings were detailed in a report published on October 21, 2026, marking a significant evolution in cybercrime tactics. The techniques employed are not new, but their integration into a self-directed agent lowers the barrier for other attackers to execute similar campaigns.