Why does my AI assistant recommend packages that turn out to be malware?

When a code-generating model scaffolds a project, it invents package names at rates between 4.6% and 6.1% even on today's frontier models. Attackers register those hallucinated names as malicious packages before developers notice. Researchers studying five frontier models released between October 2025 and March 2026 found 127 package names all five models invent identically; 53 of them remain registrable by an attacker after each registry's existing defenses. No IDE plugin, no CI step, and no package manager checks whether a model-suggested package name was ever real before resolving it. The only current advice is to read every import manually, which no team does at scale.