Category: AI Agent Security | Reading time: 3 min
Symantec published research on October 21 describing an autonomous offensive agent that its operators call "Skynet mode." The agent searches the internet for exposed AI model endpoints, takes them over, and then uses the hijacked model to run its own attack campaign without a human approving each step.
How it works
Once it controls an exposed model, the agent uses it to scan the internet, exploit web frameworks, steal databases, harvest credentials and mine cryptocurrency. Symantec says it can run for up to 100,000 steps unattended. To get the hijacked model to cooperate, the operators used a jailbreak that framed credential theft as routine development work.
Symantec captured the agent this summer through an AI honeypot, a deliberately exposed AI service set up to attract attackers.
Who pays
The cost lands on the organization whose model was hijacked. Every step the attacker's agent takes is inference the victim is billed for, and the attacks against third parties are carried out from the victim's infrastructure.
Why it matters
The pattern known as LLMjacking has usually meant stolen keys resold for free model access. This case goes further: the stolen model becomes the engine of an autonomous attack. For any company running model endpoints, that raises two practical requirements. Endpoints need authentication and network controls so they are not reachable in the first place. And when something does go wrong, the owner needs a record of what ran on its model and on whose behalf, strong enough to show regulators and affected parties that the activity was not its own.
Source: Symantec / Security.com