AI-Powered Dependency Decisions Introduce, Ignore Security Bugs

Why It Matters

Faulty AI‑driven upgrade advice inflates security risk, developer waste, and technical debt, undermining the promise of AI‑enhanced software supply‑chain management.

Key Takeaways

• Frontier AI models hallucinate 28% of dependency recommendations
• Even top models invent 1 in 16 upgrades
• Grounded AI reduces critical risks by ~70%
• Unchecked AI advice leaves 800 critical vulnerabilities
• Human‑in‑the‑loop cannot fix missing real‑time data

Pulse Analysis

The rise of large language models (LLMs) in DevSecOps promised faster, smarter decisions for software supply‑chain management. Yet, as Sonatype’s extensive study shows, the very models touted for their reasoning prowess often lack the real‑time context needed to evaluate library versions, vulnerability data, and enterprise policies. When developers query these frontier models—such as GPT‑5.2, Claude Sonnet 4.5, or Gemini 3 Pro—the output can include non‑existent versions or downgrade paths that silently retain high‑severity flaws, turning AI from a productivity boost into a hidden liability.

Sonatype analyzed 258,000 AI‑generated upgrade suggestions across the four major package ecosystems, uncovering that roughly one‑third of components received a "no change" recommendation, while the remaining advice contained a mix of hallucinations and subtle mis‑configurations. Notably, even the most advanced models still fabricated about 6% of recommendations (one in sixteen) and failed to flag 800 critical and 900 high‑severity vulnerabilities left in production. By integrating live registry data, vulnerability scores, and a version‑recommendation API—what Sonatype calls a "grounded" approach—the company achieved a near‑70% reduction in risky outcomes, demonstrating that contextual intelligence is the missing piece for trustworthy AI assistance.

For enterprises, the takeaway is clear: AI tools must be coupled with up‑to‑date dependency intelligence and policy enforcement before they can be trusted in production pipelines. Relying on human review alone cannot compensate for the lack of real‑time data, as it merely shifts the burden downstream. Organizations should adopt hybrid solutions that embed live security feeds and enforce constraints at inference time, while treating AI recommendations as advisory rather than authoritative. This strategy not only curtails technical debt but also safeguards the broader software supply chain against the inadvertent introduction of vulnerable components.