In October, Anthropic introduced an AI model called Claude Computer Use, enabling the Claude neural network to independently operate a computer based on user instructions. However, this feature, still in beta testing, has drawn attention for its potential misuse.
Cybersecurity Expert Uncovers Exploit
Cybersecurity expert Johann Rehnberger demonstrated how the Computer Use feature could be abused. In a report, he detailed how the AI downloaded and launched a malicious application at his request and subsequently communicated with the server controlling the malware.
Anthropic had warned users about potential risks associated with the feature, stating: “We recommend taking precautions to isolate Claude from sensitive data and activity to avoid risks associated with query injections.” Similar attacks targeting AI systems remain prevalent, even in early-stage testing phases.
ZombAIs and Other Vulnerabilities
Rehnberger named his exploit ZombAIs. Using this method, he forced the AI to load the Sliver remote control environment, initially designed for penetration testing but repurposed by cybercriminals. He emphasized that this is just one example of how AI can be exploited, noting that Claude could also be manipulated to write and compile malicious code in languages like C.
Other vulnerabilities have surfaced as well, notes NIX Solutions. For instance, the Chinese chatbot DeepSeek AI was found susceptible to attacks via query injections. Additionally, large language models were tricked into generating code with ANSI control characters, enabling the hacking of system terminals. This subtype of attack has been dubbed Terminal DiLLMa.
As beta testing continues, the risks associated with these advanced AI tools highlight the need for robust safeguards. Yet we’ll keep you updated as more security measures and integrations are introduced to address these concerns.