CyberSecEval 4

Meta's fourth CyberSecEval release, adding CrowdStrike-built defensive benchmarks (malware analysis, threat-intel reasoning) and an automated-patching benchmark to the prior risk and insecure-code tests.

Also known as: CYBERSECEVAL 4, Purple Llama CyberSecEval 4

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorysafety
SubcategoryLLM cybersecurity risk and capability suite: adds CrowdStrike-built defensive benchmarks (malware analysis, threat-intelligence reasoning, as CyberSOCEval) and an automated-vulnerability-patching benchmark (AutoPatchBench) to the prior version's risk and insecure-code tests
Page statusactive
Metricper-category rates and scores -- risk/insecurity rates for most inherited categories; multi-answer accuracy plus Jaccard similarity for the new CyberSOCEval tasks; patch-verification success rate for AutoPatchBench
Directionlower_is_better
Unit%
PublisherMeta (the Purple Llama project), with CyberSOCEval developed in partnership with CrowdStrike

What it measures

CyberSecEval 4 builds on CyberSecEval 3's risk categories and adds three new benchmarks the repository documents as its headline additions: two developed with CrowdStrike under the name CyberSOCEval -- Malware Analysis (can a model correctly answer multi-answer questions about a piece of malware from its detonation report) and Threat Intelligence Reasoning (can a model extract correct answers from unstructured threat-intelligence reports, as text or as report images) -- and AutoPatchBench, which measures whether an LLM agent can generate a security patch for a real, fuzzer-discovered crash in native (C/C++) code that a verification pipeline confirms actually fixes the bug. Unlike most of the suite, the CyberSOCEval benchmarks measure a defensive capability that is good to have more of, not a risk to minimize -- the first point in the CyberSecEval series where a higher score is unambiguously the desirable direction for part of the suite. CyberSecEval 4 also carries forward CyberSecEval 3's MITRE compliance, MITRE False Refusal Rate, instruct and autocomplete insecure-code tests, and multilingual prompt injection and multi-turn phishing tests, largely unchanged.

Task format

Continuing categories (largely unchanged from CyberSecEval 3, confirmed by exact sample counts): MITRE compliance (1,000 prompts), MITRE FRR (750 prompts), instruct and autocomplete insecure-code tests (1,916 prompts each), multilingual prompt injection (1,004 cases, the same file as CyberSecEval 3), and multi-turn phishing (a subset of the CyberSecEval 3 challenge set). New: Malware Analysis (609 multiple-answer questions built from detonation reports of public malware samples targeting Windows-operated businesses, with up to 10 possible correct options per question), Threat Intelligence Reasoning (588 question-answer pairs drawn from 45 distinct threat-intelligence reports, answerable from report text, report images, or both), and AutoPatchBench (an LLM agent is given a fuzzer-discovered crash from 178 bugs across 11 crash types, drawn from the ARVO dataset of reproducible open-source vulnerabilities, and must produce a patch that a fuzzing-and-differential- testing pipeline verifies actually resolves the crash without breaking the target).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub