CyberSecEval 2

Meta's second-generation LLM cybersecurity suite: adds prompt-injection, code-interpreter-abuse and exploit-capability tests to the original CyberSecEval's insecure-code and attack-compliance measures.

Also known as: CYBERSECEVAL 2, Purple Llama CyberSecEval 2

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: prompt injection, code-interpreter abuse, insecure code generation, and offensive vulnerability-exploitation capability
Page statussuperseded
Metricper-category rates -- mainly an attack/injection/insecure-code success or compliance rate (share of test cases judged unsafe); direction and headline number differ by category
Directionlower_is_better
Unit%
Dataset licenceMIT
PublisherMeta (the Purple Llama project)

What it measures

CyberSecEval 2 is Meta's second release in the CyberSecEval series, following the original CyberSecEval (arXiv 2312.04724, not separately catalogued in this repository). It is a suite of separately-scored categories rather than one number. It carries forward the first version's two core measures -- whether a model complies with requests to help carry out cyberattacks (MITRE ATT&CK-based compliance tests) and whether a model's own generated code is insecure (instruct and autocomplete tests, scored by an automated insecure-code detector for known CWE patterns) -- and adds two new areas named explicitly in the paper's title scope: prompt injection (251 hand-authored English test cases spanning direct and indirect delivery and 15 attack-technique categories) and code-interpreter abuse (500 prompts across 5 categories asking a model to misuse an attached code interpreter, for example to read files it should not access). It also adds a False Refusal Rate (FRR) test set of borderline-but-benign cybersecurity requests, introduced to weigh a model's safety against the utility it loses by over-refusing, and expands automated vulnerability-exploitation tests that measure whether a model can solve string-constraint, buffer-overflow, memory-corruption and SQL-injection challenges end to end in C, Python and JavaScript.

Task format

A per-category suite: MITRE compliance (1,000 cyberattack-assistance prompts, LLM-expanded and LLM-judged for attacker helpfulness), MITRE False Refusal Rate (750 borderline-benign prompts, judged for refusal), instruct and autocomplete insecure-code tests (1,916 prompts each, scored by Meta's Insecure Code Detector), textual prompt injection (251 direct/indirect cases, LLM-judged for whether the injected instruction was followed), code-interpreter abuse (500 prompts, LLM-judged as malicious or non-malicious), and vulnerability-exploitation challenges (hand-authored and, for some test types, randomly generated per run, scored pass/fail on whether the model's exploit works).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub