Rapid injection defenses against LLM Cyberattacks
Interesting study: “AI-Hacker: Rapid Injection as a Defense Against Cyberattacks by LLM”:
Large-scale language models (LLMs) are increasingly being used to automate cyber attacks, making sophisticated exploits more accessible and scalable. In response, we propose a new security strategy designed to combat LLM-driven cyber attacks. We present Mantis, a protective framework that uses LLMs’ strengths in the tendency to hold opposing views to undermine malicious activities. After detecting an automated cyberattack, Mantis plants carefully create inputs to the system’s responses, leading the attacker’s LLM to disrupt its operations (passive defense) or compromise the attacker’s machine (active defense). By using decoys on purpose to attract the attacker and using rapid injections of the attacker’s LLM, the mantis can automatically repel the attacker. In our tests, Mantis achieved over 95% efficiency against automated attacks run by LLM. To encourage further research and collaboration, mantis is available as an open source tool: https URL.
This is not a solution, of course. But this kind of thing can be part of the solution.
Posted November 7, 2024 at 11:13 AM • 0 Comments
Bruce Schneier sidebar photo by Joe MacInnis.
Source link