Small models help find memory safety flaws in software

A system using a 4B-parameter model found 187 memory-safety findings across 4,000 projects.

Researchers used a small decision model to find memory-safety vulnerabilities. This system searched more than 4,000 projects. It found 187 confirmed findings.

The system uses Intern-Decision-4B, which is a 4B-parameter open-weight decision model. It scores candidate functions to rank them. The process uses deterministic incomplete-fix miners and shape localizers to decide where to look.

A decision model is a model that answers a typed question in one forward pass. It returns scores for permitted answers rather than text. A harness is a surrounding system that supports the model.

The model is weak when used alone. It separates vulnerable functions from benign ones with an AUC of 0.730.73. It separates a vulnerable function from its own patch with only 0.560.56.

Recall is limited mainly by where the localizers look. The decision step runs at 4.4 decisions per second on one GB10. It runs at 9.8 decisions per second on a consumer RTX 4090 gaming PC.

The 187 findings include 12 critical, 148 high, 23 medium, and 4 low findings by worst-case CVSS v3.1. All findings are being disclosed to maintainers under coordinated disclosure.

Key facts

  • The system searched 900-1,100 repositories per hour.
  • The model placed 7.5 times more confirmed findings in the top 10 than random order.
  • The campaign used three GB10 systems.

All sections

Edition