Transfer lens
PUBLISHED METHOD / MEASUREMENT NOTE

CyberSecEval 4 · Inspect adaptation

Evaluate eight cybersecurity tasks under explicit scoring rules.

01 / TEST SETTING

Pinned datasets, task configurations and model-based or rule-based graders.

02 / OBSERVED CHECK

Task-specific scores; their meanings differ, so cross-task averages are inappropriate.

03 / TRANSFER BOUNDARY

The public implementation excludes autonomous uplift and patch validation. Reported results cover a cost-limited subset.

MEASURED TASK AREAS
AnalysisEvidence

Task classification identifies scope. It does not report model performance or safe deployment.

REPORTED FINDING

A grading denominator change breaks score comparability

Inspect documents excluding unparseable prompt-injection judgments from scoring in version 4-B, replacing their previous zero scores.

This documented scoring change affects comparability with version 3-B; it is not evidence that a model became less secure.