Security researchers have discovered an AI agent that is effectively a super cybercriminal in its own right, capable of stealing databases and credentials on a vast scale without any human assistance whatsoever. According to the Symantec Threat Research Team, the criminal AI agent's modus operandi is to illegally gain access to existing legitimate AI models and make them pay for the computer power it uses to scan the internet for targets in order to steal corporate data and credentials from a wide range of target organizations.
"Attackers have built an autonomous offensive agent that hunts the internet for exposed large language models (LLMs) and hijacks them to power a self-directed cybercrime operation," says the research team.
Once connected to an unprotected model, the agent scans for new targets, exploits common web frameworks, steals databases and credentials, and quietly installs cryptocurrency miners, conducting up to 100,000 steps without any need for human approval or assistance. The owner of the hijacked model pays for the computing power, while the servers the agent breaks into are stripped of data and put to work mining.
A fully autonomous criminal operation
What makes this tool unique is that the criminal AI agent does not simply answer its creators' questions — it plans and drives the attack entirely autonomously. To make a formerly legitimate commercial AI model cooperate with their criminal designs, the operators simply wrapped their instructions in what is known as a "jailbreak", telling the duped AI agent to ignore its safety controls and to treat credential theft as routine development work.
The "honeypot" trap
Symantec's researchers had therefore to use equal cunning in order to capture the criminal AI agent's full instruction set. They created an AI "honeypot" — an internet-facing AI service designed purely to attract and then study attacks being carried out against AI infrastructure. The criminal AI agent, believing the honeypot to be a real AI model, sent the tool's entire working prompt, thereby giving the Symantec researchers a uniquely complete view of how such an agent is built and its nefarious capabilities.
The criminal AI agent discovered by the Symantec Threat Research team may not, however, be unique, as professionally organized cybercriminal groups have been even quicker to adopt agentic AI than have legitimate enterprises. There may be other equally autonomous criminal AI agents already in the wild or under development. By corrupting otherwise legitimate AI models, a new generation of cybercriminal AI agents may now represent a rapidly growing threat for companies who have chosen to deploy their own AI agents at a time when their security is still very much in question.
