Artificial intelligence may install malicious code on your computer
Do you allow your artificial intelligence tools to write and install code independently? You should think again. A new study reveals that the most popular AI tools, including Claude from Anthropic and Codex from OpenAI, could become a conduit for injecting malware directly into your computers and corporate networks. The new risk stems from the way these tools read and execute instructions from websites.
Do you allow your artificial intelligence tools to write and install code independently? You should think again. A new study reveals that the most popular AI tools, including Claude from Anthropic and Codex from OpenAI, could become a conduit for injecting malware directly into your computers and corporate networks. The new risk stems from the way these tools read and execute instructions from websites.
How does AI install malicious code?
The problem starts with special text files called llms.txt, which are designed to help artificial intelligence agents read websites easily. Researchers who scanned over 6,000 domains of giant companies and defense contractors found that many of them contain references to code packages or domains that are not registered at all. When the AI attempts to install the software according to the erroneous documentation, it may download malicious code that hackers have registered under those same abandoned names.
To prove the danger, the researchers themselves registered some of these abandoned names and created harmless code packages that only report their installation. The results were swift and alarming: in less than an hour, a giant company from the Fortune 500 list had already contacted their server following an automatic installation performed by its AI agent. Shortly thereafter, dozens of other companies that fell into the exact same trap joined it.
How can you protect the organization?
To resolve this security breach, a dual action is required. First, companies and organizations must clean up the documentation on their websites and ensure that it does not contain links to non-existent code packages or abandoned domains. Second, AI developers must prevent agents from treating text documents as direct execution commands. Until this happens, it is recommended to limit AI permissions and not allow it to run system commands automatically.
Ultimately, artificial intelligence brings with it immense efficiency, but also new and sophisticated risks. Maintaining clean documentation and limiting the execution powers of AI agents are the most necessary steps you can take today to ensure that your digital assistant does not accidentally become the gateway for the next hacker.
