SoK: How LLM Capability Evolved in Vulnerability Research
WHY IT MATTERS
A systematization-of-knowledge paper assesses how LLM capability in software vulnerability research has evolved. Structured across the security research lifecycle.
What Happened
A systematization-of-knowledge (SoK) paper on ArXiv consolidates prior work on LLM capability in software vulnerability research, organizing findings across the security research lifecycle rather than treating model performance as a single aggregate score. The paper maps capability to discrete stages — reconnaissance, code comprehension, vulnerability identification, exploit development, and patch analysis — and assesses where current models contribute versus where they remain unreliable. It functions as a structured reference rather than a novel empirical claim, aggregating benchmarks and reported results into a lifecycle-aligned framework.
Why It Matters
Security tooling teams have lacked a shared vocabulary for evaluating where LLMs actually help versus where benchmark scores overstate capability. This SoK supplies that scaffolding, letting builders position their products against lifecycle stages instead of competing on generic "vulnerability detection" claims. The practical benefit is procurement and roadmap clarity: a team can now ask which stage a model or pipeline targets, what evidence supports that claim, and where the failure modes concentrate. For operators running detection or triage workflows, the framework surfaces the gap between identification and exploitation-adjacent reasoning, which is where deployment risk and false confidence tend to accumulate. It also gives evaluation teams a defensible structure for internal benchmarking that maps to how security work is actually sequenced.
Technical Details
The paper segments capability assessment across the vulnerability research pipeline, treating code comprehension and vulnerability identification as distinct from exploit generation and patch validation. Reported performance varies materially by stage: identification tasks benefit from retrieval augmentation and larger context windows, while exploit development remains dependent on execution feedback loops and environment access rather than static reasoning alone. Benchmark coverage is uneven — the paper notes that identification benchmarks are comparatively mature while end-to-end exploit and patch-analysis evaluations are sparse and often non-reproducible. Limitations flagged include benchmark contamination risk, weak generalization across codebases and languages, and the mismatch between curated CVE-derived datasets and real-world repository noise. Integration requirements center on grounding mechanisms — retrieval, static analysis scaffolding, sandboxed execution — since unassisted model output degrades sharply outside the identification stage.
Operational Impact
Teams building security tooling can restructure evaluation suites around lifecycle stages, replacing single aggregate scores with per-stage pass rates that map to customer workflows. Detection and triage pipelines become the near-term deployment target, since that is where the evidence base is strongest; exploit-adjacent automation stays gated behind human review and sandboxed execution. Benchmark contamination awareness changes how operators select evaluation corpora — favoring held-out and time-split data over public CVE sets. The framework also reduces duplicated effort: instead of rebuilding stage taxonomies internally, teams adopt a shared reference and spend engineering time on grounding infrastructure. Patch-analysis workflows, currently under-instrumented, become a visible gap to prioritize.
SOURCE
ArXiv
SHARE
MORE FROM STUFFINSIDER
Depth Pro 2: Sharp, Temporally Consistent Monocular Depth Estimation
Nov 20RESEARCHRL Post-Training in Diffusion Transformers Works via Activation Steering
Nov 20RESEARCHOpenAI Releases MedChemBench Benchmark for Drug Candidate Identification
Nov 21RESEARCHOpenAI Research: Defending Models Against Adversarial Distillation
Nov 21