Pervasive LLM Hallucinations Expand Code Developer Attack Surface

Lasso Security research found that major LLMs (GPT-3.5, GPT-4, Gemini Pro, Coral/Cohere) often invent nonexistent package names when asked for coding libraries; attackers can exploit this by publishing malicious packages with those exact names—one such proof-of-concept upload of a hallucinated "huggingface-cli" package received more than 32,000 downloads. The study measured high hallucination and repetition rates (e.g., Gemini ~64.5% hallucination rate; repetitiveness up to ~24% for some models) and warns developers to verify package provenance, community activity, maintenance, and scan packages before adding them to development environments.