Top AI tools such as OpenClaw and Github Copilot can be hijacked to create new massive botnets (www.techradar.com)

Researchers from Intuit, Technion, and Tel Aviv University have unveiled a new security vulnerability known as HalluSquatting, which exploits the AI phenomenon of hallucination in popular tools like GitHub Copilot and OpenClaw. This technique allows attackers to hijack these AI services to deploy malicious code, effectively transforming devices into botnets capable of conducting large-scale ransomware attacks. HalluSquatting draws parallels to typosquatting but focuses on the inaccuracies in Large Language Models (LLMs) when identifying resources. By preemptively registering fake resources that an LLM is likely to misidentify, hackers can execute remote code attacks simply by having unwitting users trigger the faulty resource. This security revelation raises significant concerns for the AI/ML community as the potential for large-scale exploitation of LLM-based platforms increases. Previous malware strategies have utilized pull-based attacks, but HalluSquatting combines push and pull methodologies for greater impact. While some mitigation strategies, like robust resource naming conventions and limiting LLM fetch operations, could help address these vulnerabilities, the diverse nature of AI tool developers means effective implementation could be slow. As the frequency of LLM-related attacks rises, the community must prioritize security measures to safeguard against these emerging threats.