Giant Language Fashions (LLMs) have a critical “bundle hallucination” downside that would result in a wave of maliciously-coded packages within the provide chain, researchers have found in one of many largest and most in-depth ever research to research the issue.
It’s so dangerous, actually, that throughout 30 totally different exams, the researchers discovered that 440,445 (19.7%) of two.23 million code samples they generated experimentally in two of the most well-liked programming languages, Python and JavaScript, utilizing 16 totally different LLM fashions for Python and 14 fashions for JavaScript, contained references to packages that had been hallucinated.
The multi-university examine, first revealed in June however lately up to date, additionally generated “a staggering 205,474 distinctive examples of hallucinated bundle names, additional underscoring the severity and pervasiveness of this menace.”