| --- |
| extra_gated_prompt: >- |
| SecProbe is in a private preview. Access is granted manually by the authors to invited collaborators. |
| The bundles contain hidden grader tests and reference solutions: do not train on them or redistribute them. |
| extra_gated_fields: |
| GitHub username: text |
| Affiliation: text |
| I will not train models on this dataset or redistribute its hidden tests or solutions: checkbox |
| extra_gated_button_content: Request access |
| license: other |
| pretty_name: SecProbe |
| task_categories: |
| - text-generation |
| tags: |
| - security |
| - code |
| - agents |
| - benchmark |
| size_categories: |
| - n<1K |
| --- |
| |
| # SecProbe |
|
|
| SecProbe is a benchmark of **353 repository-scale vulnerability-repair tasks**. Each task is a working |
| software project (Python, Go, C++, TypeScript, Java or Rust) into which several plausible, CWE-specific |
| security defects were seeded. The agent must audit the repository, patch the root causes without breaking |
| documented behaviour, and is graded by hidden functional and adversarial tests (score 0-100; a task is |
| passed only at 100). |
|
|
| > **Do not train on this dataset.** Every bundle contains the hidden grader tests (`tests/`) and the |
| > reference repair (`solution/`). Using it for training or fine-tuning contaminates the benchmark. |
| > Please also avoid publishing the hidden tests or solutions elsewhere. |
|
|
| ## Layout |
|
|
| ``` |
| index.json # task index (id, CWE, category, language, level, sizes, ...) |
| tasks/<task_id>.tar.gz # one Harbor task per archive; extracts to <task_id>/ |
| ``` |
|
|
| Each archive extracts to a [Harbor](https://github.com/laude-institute/harbor) task directory: |
|
|
| ``` |
| <task_id>/ |
| task.toml # metadata, timeouts, resources |
| instruction.md # the prompt given to the agent |
| environment/ # Dockerfile + the vulnerable repository (copied to /app) |
| tests/ # hidden verifier (grader + hidden tests) <- do not show to agents |
| solution/ # reference repair <- do not show to agents |
| ``` |
|
|
| ## Running one task |
|
|
| ```bash |
| TASK=patch-cwe94-0012 |
| curl -L -o $TASK.tar.gz https://huggingface.co/datasets/<HF_DATASET>/resolve/main/tasks/$TASK.tar.gz |
| tar xzf $TASK.tar.gz |
| harbor run -p $TASK -a terminus-2 -m <model> |
| ``` |
|
|
| `<model>` is a LiteLLM-style `provider/model` id (e.g. `anthropic/claude-opus-4-8`). Any Harbor agent |
| (`terminus-2`, `mini-swe-agent`, `claude-code`, `codex`, `openhands`, ...) can be used with `-a`. |
| The verifier writes a 0/1 reward (pass = score 100) and the 0-100 score. |
|
|
| ## Citation |
|
|
| Please cite the SecProbe paper if you use this benchmark. |
|
|
| ## Links |
|
|
| - Paper: https://arxiv.org/abs/2609.33763 |
| - Project page (browse every task, evaluate agents, generate new tasks): https://xiaonan-debug.github.io/secprobe-page/ |
|
|
