subtle-ways-to-tell-a-lie

Subtle Ways to Tell a Lie: Slopsquatting and Supply-Chain Security

By Angelica, Philippe

Threat Model Summary

An attacker registers a package name commonly hallucinated by AI code assistants, delivering malicious code to developers who execute copy-pasted recommendations.

Key Findings

  • AI code assistants consistently suggest plausible but non-existent package names when prompted for trendy or popular functionality.
  • Monitoring and pre-emptively registering these hallucinated package names (slopsquatting) allows attackers to execute targeted supply chain attacks.
  • Standard package registries lack validation for whether suggested imports are safe or verified, increasing importance of client-side guardrails.

Models Tested

qwen2.5-coderdeepseek-codercodellamaqwen2.5llama3.1

Taxonomy Tags

LLM01LLM02T1195.001

This report investigates the pre-conditions of slopsquatting: a supply-chain attack vector born from the systematic tendencies of Large Language Models to hallucinate non-existent software packages during code generation. Rather than relying on a hostile data path into an active system, this vulnerability exploits routine development workflows, becoming particularly pronounced when developers use trend-seeking language to prompt models for modern, popular, or recent ecosystem tools.