- Researchers redefined Eco -Fax Hack and saw that AI do everything without direct control
- AI model successfully violated a huge violation with zero human input
- Shell’s orders did not require, AI worked as a planner and assigned everything else
Large language models (LLM) have long been considered to be useful tools in areas such as data analysis, content production, and code aid.
However, a new research by Carnegie Milen University, which was done in partnership with Anthropic, has raised difficult questions about his role in CyberScript.
This research shows that in the right circumstances, LLMS can plan and carry out complex cyberrtex without human guidance, which can only suggest a change in complete sovereignty with digital interference.
You can like
From puzzles to enterprise environment
The first experiences with AI in CyberScureti were mostly limited to the “Flag on Capture Di” scenarios, easy challenges used for training.
The Carnegie Millen team, headed by PhD candidate Brian Singer, proceeded further by giving structural guidance to the LLM and integrated into the agents rankings.
With these settings, they managed to test the model in a more realistic network setup.
In one case, they re -made the same conditions that violated the 2017 Eco -Fax, which included documentary threats and setting in official reports.
The AI not only planned the attack but also deployed malware and extracted data directly, without human orders.
What makes this research amazing is how little the LLM had to do. Traditional approaches often fail because the model struggles for a detailed login to the model shell commands.
Instead, the system relied on a high -level structure where the LLM worked as a planner, assigning lower -level steps to all agents.
This abstract provided enough context to “understand” AI and adapt to its environment.
Although these results were achieved in the order of a control lab, they raise questions about what extent this sovereignty can go.
The dangers here are not just a fictitious concept. If the LLMS itself can violate the network, malicious actors can use them to measure more attacks than they are possible with human teams.
Even tools such as end -point protection and excellent anti -virus software can be tested by such adaptive and responsible agents.
Nevertheless, this ability has potential benefits. A LLM system capable of imitating realistic attacks can be used to improve and expose flaws that otherwise no one is careful.
The singer explained, “It only operates under specific terms, and we have nothing that can attack the Internet independently … but this is an important first step.”
Nevertheless, AI’s ability should not be excluded to develop a major violation with at least input.
Follow -up research is now finding out how these techniques can be implemented in defense, potentially enables AI agents to detect or prevent attacks in real time.
