Adaptive textual payloads for disrupting LLM-powered cyberattacks
A system and method for defending against cyberattacks executed by one or more large language model (LLM)-powered agents presented. The method includes deploying one or more traps throughout a computing environment to protect at least one protected entity, wherein each trap of the one or more traps is designed to exploit one or more vulnerabilities of a LLM contained within an LLM-powered agent; monitoring, in real-time, a plurality of interactions with the deployed traps; detecting, based on the monitored plurality of interactions, interaction behavior consistent with that of at least one LLM-powered agent of the one or more LLM-powered agents; and responsive to the detection, causing one or more of the deployed traps to initiate one or more defensive actions.
1 . A method for defending against cyberattacks executed by one or more large language model (LLM)-powered agents, comprising:
deploying one or more traps throughout a computing environment to protect at least one protected entity, wherein each trap of the one or more traps is designed to exploit one or more vulnerabilities of a LLM contained within an LLM-powered agent;
monitoring, in real-time, a plurality of interactions with the deployed traps;
detecting, based on the monitored plurality of interactions, interaction behavior consistent with that of at least one LLM-powered agent of the one or more LLM-powered agents; and
responsive to the detection, causing one or more of the deployed traps to initiate one or more defensive actions.
2 . The method of claim 1 , wherein a defensive action of the one or more defensive actions comprises performance of any one of: honeying, cloaking, or trapping.
3 . The method of claim 2 , wherein the defensive action is honeying, and wherein the method further comprising:
deploying any one of: at least one LLM-specific honeytoken and, at least one LLM-specific honeypot, or both.
4 . The method of claim 3 , wherein detecting the interaction behavior further comprises:
obtaining a detection signal identifying a presence of the at least one LLM-powered agent using the at least one LLM-specific honeytoken, the at least one LLM-specific honeypot, or a combination thereof.
5 . The method of claim 2 , wherein the defensive action is cloaking, and wherein the method further comprising:
delaying the at least one LLM-powered agent.
6 . The method of claim 5 , further comprising:
diverting attention of the at least one LLM-powered agent away from the at least one protected entity.
7 . The method of claim 2 , wherein the defensive action is trapping, and wherein the method further comprising:
stopping the at least one LLM-powered agent.
8 . The method of claim 2 , wherein causing the deployed traps to initiate the one or more defensive actions further comprises:
generating an alert on a cyberattack executed by the at least one LLM-powered agent.
9 . The method of claim 1 , wherein deploying the one or more traps throughout the computing environment further comprises:
embedding the traps within the at least one protected entity.
10 . The method of claim 9 , wherein the one or more deployed traps are programmed to: modify existing strings, create new strings, or a combination thereof.
11 . The method of claim 1 , wherein monitoring, in real-time, the plurality of interactions with the deployed traps further comprises:
logging one or more requests to a planted URL and correlating the requests to a tokenized target.
12 . The method of claim 1 , wherein one or more of the one or more traps are programmed to present content that is parsed and/or interpreted distinctly from at least one of: a LLM, tools, a framework, and a knowledge base of the at least one LLM-powered agent.
13 . The method of claim 1 , wherein the at least one LLM-powered agent includes any one of an autonomous pentesting agent, a semi-autonomous pentesting agent, an artificial intelligence pentesting agent, and a human-operated pentesting agent.
14 . A non-transitory computer-readable medium storing a set of instructions for defending against cyberattacks executed by one or more large language model (LLM)-powered agents, the set of instructions comprising:
one or more instructions that, when executed by one or more processors of a device, cause the device to:
deploy one or more traps throughout a computing environment to protect at least one protected entity, wherein each trap of the one or more traps is designed to exploit one or more vulnerabilities of a LLM contained within an LLM-powered agent;
monitor, in real-time, a plurality of interactions with the deployed traps;
detect, based on the monitored plurality of interactions, interaction behavior consistent with that of at least one LLM-powered agent of the one or more LLM-powered agents; and
responsive to the detection, cause one or more of the deployed traps to initiate one or more defensive actions.
15 . A system for defending against cyberattacks executed by one or more large language model (LLM)-powered agents comprising:
one or more processors configured to:
deploy one or more traps throughout a computing environment to protect at least one protected entity, wherein each trap of the one or more traps is designed to exploit one or more vulnerabilities of a LLM contained within an LLM-powered agent;
monitor, in real-time, a plurality of interactions with the deployed traps;
detect, based on the monitored plurality of interactions, interaction behavior consistent with that of at least one LLM-powered agent of the one or more LLM-powered agents; and
responsive to the detection, cause one or more of the deployed traps to initiate one or more defensive actions.
16 . The system of claim 15 , wherein a defensive action of the one or more defensive actions comprises performance of any one of:
honeying, cloaking, or trapping.
17 . The system of claim 16 , wherein the one or more processors, when the defensive action is honeying, and are configured to:
deploy any one of:
at least one LLM-specific honeytoken and, at least one LLM-specific honeypot, or both.
18 . The system of claim 17 , wherein the one or more processors, when detecting the interaction behavior, are configured to:
obtain a detection signal identifying a presence of the at least one LLM-powered agent using the at least one LLM-specific honeytoken, the at least one LLM-specific honeypot, or a combination thereof.
19 . The system of claim 16 , wherein the one or more processors, when the defensive action is cloaking, and, are configured to:
delay the at least one LLM-powered agent.
20 . The system of claim 19 , wherein the one or more processors are further configured to:
divert attention of the at least one LLM-powered agent away from the at least one protected entity.
21 . The system of claim 16 , wherein the one or more processors, when the defensive action is trapping, and, are configured to:
stop the at least one LLM-powered agent.
22 . The system of claim 16 , wherein the one or more processors, when causing the deployed traps to initiate the one or more defensive actions, are configured to:
generate an alert on a cyberattack executed by the at least one LLM-powered agent.
23 . The system of claim 15 , wherein the one or more processors, when deploying the one or more traps throughout the computing environment, are configured to:
embed the traps within the at least one protected entity.
24 . The system of claim 23 , wherein the one or more deployed traps are programmed to:
modify existing strings, create new strings, or a combination thereof.
25 . The system of claim 15 , wherein the one or more processors, when monitoring, in real-time, the plurality of interactions with the deployed traps, are configured to:
log one or more requests to a planted URL and correlating the requests to a tokenized target.
26 . The system of claim 15 , wherein one or more of the one or more traps are programmed to present content that is parsed and/or interpreted distinctly from at least one of:
a LLM, tools, a framework, and a knowledge base of the at least one LLM-powered agent.
27 . The system of claim 15 , wherein the at least one LLM-powered agent includes any one of an autonomous pentesting agent, a semi-autonomous pentesting agent, an artificial intelligence pentesting agent, and a human-operated pentesting agent.