Symantec discovers autonomous AI agent capable of executing cyberattacks

Symantec researchers have identified an autonomous artificial intelligence agent capable of planning and executing digital invasions without human intervention. The tool utilizes a technique known as ‘LLMjacking,’ which involves the unauthorized use of large language models (LLMs) exposed on the internet to hijack computational power from third parties.

By targeting AI structures, gateways, and services connected to the Model Context Protocol (MCP), the agent shifts processing costs to victims while reducing the risk of detection. Through an AI honeypot experiment that simulated an unprotected service, researchers recorded over 127,000 external requests from tens of thousands of IP addresses. This allowed the team to intercept the agent's instructions and reconstruct its operational methods.

The intercepted commands revealed approximately 74 offensive capabilities, including botnet control, credential theft, data extraction, and network scanning. To bypass the protections of commercial models, the agent employs layered instructions that disguise malicious actions—such as installing Monero miners or hiding files—as legitimate software development tasks. The primary innovation noted by Symantec is the integration of these diverse functions into a single agent capable of coordinating complex operations independently.