Generative AI models hallucinated non-existent software package names which were then exploited by security researchers who created real packages with those names, leading to thousands of downloads by developers following AI recommendations.
Multiple generative AI models including GPT-3.5-Turbo, GPT-4, Gemini Pro, and Command were found to hallucinate non-existent software package names when providing coding assistance. Security researcher Bar Lanyado from Lasso Security demonstrated this vulnerability by creating a fake Python package called 'huggingface-cli' after observing AI models repeatedly recommend it. The package was uploaded to PyPI in December and by February, major companies including Alibaba had incorporated installation instructions for this fake package into their GraphTranslator project documentation. Research studies found that commercial AI models hallucinate package names about 5.2% of the time while open source models do so 21.7% of the time. In one comprehensive study of 16 LLMs generating 576,000 code samples, researchers identified 205,474 unique hallucinated package names. The fake 'huggingface-cli' package created by Lanyado was downloaded thousands of times by developers following AI recommendations. While Lanyado's package was benign, the incident demonstrates how malicious actors could exploit this vulnerability by creating malware-laden packages using AI-hallucinated names, a technique dubbed 'slopsquatting' by security experts.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that inadvertently generate or spread incorrect or deceptive information, which can lead to inaccurate beliefs in users and undermine their autonomy. Humans that make decisions based on false beliefs can experience physical, emotional or material harms
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed