Hackers are using exposed, internet-accessible AI models as part of automated cyberattacks, according to Symantec research described by multiple outlets. The reported approach involves an “AI agent” that finds reachable AI models and then uses them to help plan break-ins, while the owner of the model is said to bear the computing costs.
Help Net Security, Infosecurity Magazine, and The Wall Street Journal focus on Symantec’s description of the agent’s ability to operate with little or no human intervention. Help Net Security adds detail that researchers captured the agent’s instruction set after setting up a honeypot designed to look like an unprotected AI service, and that the operators reportedly sent their working prompt to the decoy.
Bloomberg highlights the technique as a way to subvert AI safety guardrails, while the others emphasize the hijacking and autonomous attack capability. Across coverage, the common thread is that legitimate AI models exposed on the internet can be repurposed into components of cybercrime.
