AI agent designed to hijack exposed models for autonomous cyberattacks
Threat actors designed an AI agent to hijack exposed AI models and use them to run cyberattacks with no human approval for up to 100,000 steps. Symantec researchers captured its instructions in a honeypot, revealing how it was built and the range of actions it was designed to perform.
The agent was instructed to scan the internet for unauthenticated AI endpoints or ones accessible with stolen keys, then exploit common web frameworks. Its 74 offensive tools included capabilities for stealing databases, credentials and secrets, running code remotely and mining cryptocurrency; a “Skynet mode” loop was designed to find, infect and revisit targets. Symantec said it had no evidence the agent was used in real-world attacks, though its history showed activity against a live target.
In short
- Attack agent was designed to run up to 100,000 steps unattended.
- It targeted exposed AI endpoints and vulnerable web frameworks.
- Researchers found no evidence of real-world attacks.
Explained simply
AI models can be made available through online services called endpoints, which other software uses to send them requests. If these services lack access controls, or attackers obtain stolen access keys, they may be able to use them without permission.
Symantec is a cybersecurity company that investigates threats, including by monitoring decoy systems called honeypots. It said the instructions it captured described an AI agent with tools for compromising computer systems and repeating tasks automatically, but found no evidence it had been used in real-world attacks.
The case matters because it raises questions about how exposed AI services could be misused and how much oversight is needed when software can take many actions on its own. The reported capabilities show what the agent was designed to do; they do not establish that those actions were carried out at scale.
Read the full article at the source →
Originally published by The Register as “Autonomous attack agent in ‘Skynet mode’ hijacks other models to carry out its evil deeds”.