An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks
decoder ·
Symantec researchers have identified a malicious AI agent that hijacks unprotected language models to launch self-sustaining cyberattacks, a tactic they term "LLMjacking." Once access is gained, the agent bypasses safety guardrails using a complex five-layer jailbreak and operates autonomously for up to 100,000 steps in what operators call "Skynet mode." During this process, the compromised model scans for additional targets, exploits server vulnerabilities, steals credentials, and installs cryptominers without human intervention. The owner of the hijacked AI system bears the significant compute costs associated with these operations. The attack was exposed when the threat actors sent their full working prompt to a Symantec honeypot, revealing the agent's architecture and toolset. This incident highlights a growing trend where attackers integrate AI directly into offensive tools to automate large-scale intrusions and data theft. Experts warn that as autonomous agents become more capable, both intentional misuse and accidental hacking incidents are likely to increase across various sectors.
decoder