AI Code Hallucinations Increase the Risk of ‘Package Confusion’ Attacks

AI created computer code Referring to third-party libraries, gold programs for supply chain attacks, the harmful packages that can steal data are vomant programs and other newly published research.

The work that used 16 of the most commonly used large language models to create a code sample of 576,000 was the “halilination” of 440,000 packages. The most open source models are in the most, 21 percent of the dependencies connected to existing libraries. Addiction is an important code component that a separate code piece requires the correct functioning. Addictions are the obstacle of re-writing code to developers and is an important part of the modern software supply chain.

Package hallucinated flashbacks

These non-existent addictions threaten the supply chain of software by sharpening addictive confusion attacks. These attacks work, the wrong component dependence, publish a harmful package and provides the same name to the same name, but by the subsequent version of the program, including the same name. The program, which is dependent on the package, in some cases, because the first, because it looks like the previous one, will choose a malicious version, not a legal version.

It is also known as a package mixture, it was a form of attack first demonstrated In 2021, using a fake code in the networks belonging to the planet, Apple, Microsoft and Tesla’s largest companies. One type of technique that provides software is a program that aims to poison a program to infect all users down.

“The attacker has published a package under a slingshot of several harmful code, and the student and the lead researcher, if a user trusts in the exit of LLM, the attacker’s truck will be executed in the truck.”

In AI, an LLM occurs when it produces a conclusion that is not in virtually incorrect, nonsensical, nonsensical or completely unrelated. Hallucinations have a long dog LLMS because they have condemned and predict and predict and eliminate their usefulness and reliability because they have a difficult impairment. One paper 2025 Dubbes the phenomenon of “package hallucination” planned to be presented at the Usenix Security Symposium.

Researchers for research, 16 and 14 in 16 pyon programming in 16 and 14 in JavaScript, it creates a 19,200 code sample for a total of 576,000 code samples. In these samples, 2.23 million packages, 440,445 or 19.7 percent, pointed to the existing packages. Among these 440,445 package hallucination, there were unique package names between 205,474.

One of the package hallucinations, which is potentially useful in the supply chain attacks, 43 percent of package hallucinations have repeated more than 10 surveys. “In addition, the researchers” 58% of the name, most of the hallucinations are more randomly, but more than once more valuable for harmful actors who make the vector of the hallucination and the hallucination attack. “