AI Security

The Reproducibility Crisis In AI Security Evals

ML research has a reproducibility crisis. AI security evaluation inherits it. Vendors publishing numbers that can't be reproduced are the norm — not the exception.

Shadab Khan
Security Engineer
2 min read

Academic ML has a well-documented reproducibility crisis. AI-for-security evaluation inherits the problem and adds its own. Vendors quote accuracy numbers that customers cannot reproduce. Dataset details are private. Methodology documents are marketing. The pattern is common enough to call a crisis in the narrow sense — the ability to independently verify claims is rare. Customers adapting to this reality have specific evaluation practices.

What reproducibility requires

Five components:

  • Dataset with provenance.
  • Methodology document.
  • Version pinning of model and tooling.
  • Scoring code that can be rerun.
  • Variance reporting across runs.

Missing any of these breaks reproducibility.

What vendor numbers usually provide

Three components out of five:

  • Dataset: usually private.
  • Methodology: often partial.
  • Version pinning: sometimes missing.
  • Scoring code: rarely shared.
  • Variance: frequently absent.

The gap is the reproducibility gap.

What to demand

Three asks for any AI-for-security vendor:

  1. Dataset description with construction methodology.
  2. Scoring code or pseudocode.
  3. Variance numbers across repeated runs.

If the vendor can't provide all three, the numbers they quote are less reliable than they appear.

How Safeguard Helps

Safeguard publishes all five reproducibility components for its eval benchmarks. Customers can rerun key benchmarks in their own environment. For organisations whose procurement process requires defensible benchmark numbers, reproducibility is the baseline — not the bonus.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.