AI Code Tools Widely Hallucinate Packages

Researchers analyzed 16 code-generating LLMs and found a widespread problem of "package hallucinations," where models invent non-existent npm/PyPI package names; these hallucinations occur frequently (open-source models ~21.7%, commercial ~5.2%) and are often repeated across runs. The study warns that attackers can exploit repeated hallucinated names by publishing malicious packages with those names (a supply-chain attack vector labeled "slopsquatting"), shows models sometimes self-identify hallucinations, and cites prior work demonstrating real-world downloads of an uploaded invented package.