Tech Radar
Claude AI and different techniques could possibly be weak to worrying command immediate injection assaults
Dec
- Safety researchers tricked Anthropic’s Claude Pc Use to obtain and run malware
- They are saying that different AI tools could possibly be tricked with immediate injection, too
- GenAI could be tricked to jot down, compile, and run malware, as nicely
In mid-October 2024, Anthropic launched Claude Computer Use, an Synthetic Intelligence (AI) mannequin permitting Claude to manage a tool – and researchers have already discovered a strategy to abuse it.
Cybersecurity researcher Johann Rehnberger just lately described how he was in a position to abuse Pc Use and get the AI to obtain and run malware, in addition to get it to speak with its C2 infrastructure, all by means of prompts.
Whereas it sounds devastating, there are some things price mentioning right here: Claude Pc Use continues to be in beta, and the corporate did go away a disclaimer saying that Pc Use won’t all the time behave as supposed: “We recommend taking precautions to isolate Claude from delicate information and actions to keep away from dangers associated to immediate injection.” One other factor price noting is that it is a immediate injection assault, pretty widespread towards AI instruments.
“Numerous methods” to abuse AI
Rehnberger calls his exploit ZombAIs, and says he was in a position to get the device to obtain Sliver, a professional open supply command-and-control (C2) framework developed by BishopFox for crimson teaming and penetration testing, however it’s usually misused by cybercriminals as malware.
Risk actors use Sliver to ascertain persistent access to compromised techniques, execute instructions, and handle assaults in the same strategy to different C2 frameworks like Cobalt Strike.
Rehnberger additionally confused that this isn’t the one strategy to abuse generative AI instruments, and compromise endpoints through immediate injection.
“There are numerous others, like one other manner is to have Claude write the malware from scratch and compile it,” he stated. “Sure, it could possibly write C code, compile and run it.”
“There are numerous different choices.”
In its writeup, The Hacker News added DeepSeek AI chatbot was additionally discovered weak to a immediate injection assault that would enable risk actors to take over sufferer computer systems. Moreover, Giant Language Fashions (LLM) can output ANSI escape code, which can be utilized to hijack system terminals through immediate injection, in an assault dubbed Terminal DiLLMa.
You may additionally like
Source link
Copyright 2024 © Skull Sales
