license: apache-2.0
task_categories:
- question-answering
- text-generation
language:
- en
tags:
- cybersecurity
- security
- code-analysis
- vulnerability-analysis
- benchmark
- evaluation
pretty_name: FBE and VAB security analysis benchmarks
size_categories:
- n<1K
configs:
- config_name: fbe
data_files: fbe.jsonl
- config_name: vab
data_files: vab.jsonl
FBE and VAB
Two small benchmarks for security code analysis. Both grade without an LLM judge, so runs are cheap and repeatable.
FBE (find-the-bug)
14 code snippets, each with one planted vulnerability. Ask the model to analyze the code, then check whether it actually found the flaw.
Grading uses concept groups: the answer has to contain at least one synonym from every required group. Four numbers come out:
- found, did it identify the real vulnerability (this is the one that matters)
- capability, did it say what the attacker gains
- fix, did it propose a remediation at a sensible layer
- distracted, did it wander onto something irrelevant (lower is better)
Bug classes: missing object-level authorization on a sibling route, mass assignment of a role, second-order SSRF through a stored URL, path traversal where validation runs before decoding, edit-after-approval, a single-use token that is never consumed, a check-then-act race on inventory, identity taken from a client-supplied header, a worker trusting mutable queue fields, negative-amount transfer, authorization that fails open, non-constant-time secret comparison, a support route that bypasses the refund workflow, and a delete route that authenticates but doesn't authorize.
Row format:
{"id": "...", "vuln": "...", "domain": "...", "code": "...",
"must_hit": [["synonym", "..."]], "capability": [[...]], "fix": [[...]], "distractors": [[...]]}
VAB (multiple choice)
20 MCQ items over code and scenarios: broken invariant, root cause, capability (knowledge, influence, access, authority, execution), chain validity, missing check, confidence, negative space, state sequence. For chain validity the options are fixed: A proven, B probable, C speculative, D blocked.
Worth knowing before you use it: VAB rewards short forced-choice answers. A model tuned to write long-form analysis can drop on VAB while being unchanged or better at actually finding bugs. We measured exactly that. Treat FBE as the real signal and VAB as a secondary one.
Reference numbers
Qwythos / Qwen3.5-9B 4-bit base: FBE found 79%, FBE full credit 43%, VAB 75%.
Use
Authorized security evaluation and research. All snippets are generic.