Cybersecurity model scores vary widely with evaluation methods
Benchmark Scores Are Pipeline-Dependent: A Reliability Audit of Cybersecurity LLM Benchmarks
Cryptography and SecurityArtificial IntelligenceComputation and Language
Summary
When testing cybersecurity language models, the way tests are set up can drastically change the results. The authors found that changing evaluation choices can shift a model’s score by over 80 points and reorder model rankings. Even tasks that seem similar can rank models differently if evaluated differently. Their work shows the importance of carefully standardizing how these tests are done to get reliable comparisons.
What this means in practice
- •For cybersecurity ai developers: Identify and fix evaluation inconsistencies to fairly compare cybersecurity language models.
- •For ai benchmarking teams: Standardize evaluation pipelines to produce more dependable scores for language models.
Authors
Aymene Berriche, Cathrine Shalby, Mohannad Alhanahnah, Yazan Boshmaf
Abstract
Large language model (LLM) benchmarks are often treated as fixed datasets with stable scores, yet their outcomes depend on configurable evaluation pipelines. We audit eight cybersecurity benchmarks across 10 proprietary, open-weight, and cybersecurity-specialized LLMs. By modeling benchmarks as measurement pipelines, we identify 15 systematic failure modes and show that a single pipeline choice can change a model's score by more than 80 percentage points and substantially alter model rankings. At the cross-benchmark level, two semantically similar task pairs rank the same models differently because of incompatible evaluation conventions. Under an evaluation harness that standardizes pipeline choices while preserving task semantics, nine of 10 models shift by at least three ranks on at least one benchmark. These results show that cybersecurity LLM benchmark scores are pipeline-dependent and motivate pipeline-aware auditing as a core requirement for reliable model evaluation.