File size: 4,046 Bytes
1316a2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- cybersecurity
- vulnerability
- php
- benchmark
- cybergym
---

# Cybergym-Pro — PHP benchmark tasks

English | [中文](#中文说明)

Snapshot: **2026-11-04**. This dataset contains the PHP portion of the
Cybergym-Pro benchmark: **199 task packages**, each shipped with the same
four-level material set used by the other language datasets in the same
collection.

## Layout

```text
tasks/<task_id>/
  images/<task_id>-vul.tar.gz     # docker image archive, vulnerable build
  images/<task_id>-fix.tar.gz     # docker image archive, fixed build
  levels/level0/                  # repo-vul.tar.gz + README.md + submit.sh
  levels/level1/                  # + description.txt
  levels/level2/                  # + error.txt (reference input output on vul)
  levels/level3/                  # + patch.diff and repo-fix.tar.gz
  source/repo-vul.tar.gz          # vulnerable source snapshot
  source/repo-fix.tar.gz          # fixed source snapshot
  source/patch.diff               # upstream fix diff
  source/source-hashes.json       # archive hashes
  metadata.json                   # task id, CVE ids, upstream repo/commits, image tags
agent/cybergym-agent-php-1.tar.gz # PHP agent runtime image (docker save, gzipped)
```

`<task_id>` is the 16-hex Cybergym task id (e.g. `019e0f0b88563440`); the CVE
mapping is in each task's `metadata.json` (`cve_ids`).

## Level semantics

- `level0`: vulnerable source + `README.md` + `submit.sh` (run a PoC against the
  vulnerable build).
- `level1`: adds `description.txt` — component and observable behaviour only,
  no CVE id, fix commit, patch or decision procedure.
- `level2`: adds `error.txt`, the real output of the reference input on the
  vulnerable program.
- `level3`: adds `patch.diff` and the fixed source archive.

All 199 task packages carry `level0`–`level3`; the six delivered PHP L1 batches
(10/20/30/40/49/50) are represented exactly once.

## Using the images

```bash
docker load -i tasks/<task_id>/images/<task_id>-vul.tar.gz
docker load -i tasks/<task_id>/images/<task_id>-fix.tar.gz
docker run --rm --network none -v /path/to/poc:/tmp/poc:ro \
  cybergym/php:<task_id>-vul /usr/local/bin/run_poc /tmp/poc
```

Each task exposes the same runtime contract: `/usr/local/bin/run_poc <poc>`
executes the submitted PoC inside the task image. Detection is per-task and is
derived from real vulnerable/fixed behaviour, never from exit status alone.

## Agent runtime

`agent/cybergym-agent-php-1.tar.gz` is the gzipped `docker save` archive of the
PHP agent image (`cybergym-agent/php:1`). It bundles PHP 8.4, Composer, Python,
Node.js, Codex CLI, Claude Code and OpenHands.

```bash
gunzip -c agent/cybergym-agent-php-1.tar.gz | docker load
docker run --rm -it cybergym-agent/php:1
```

## Notes

- Reference PoCs, evaluation contracts and admin validation material are kept
  private and are **not** part of this dataset.
- `metadata.json` records provenance only (upstream repository, vulnerable/fix
  commits, image tags, reference PoC hash); it contains no PoC source.
- Task packages are provided for benchmark evaluation. Use them only in
  sandboxed environments you control.

## 中文说明

本数据集是 Cybergym-Pro 的 PHP 部分,共 **199** 个任务包,结构与同一 collection
中的其他语言数据集一致:

- `tasks/<task_id>/`:任务包,含漏洞/修复两份镜像归档(`images/`)、四层材料
  (`levels/level0`–`level3`)、源码快照与补丁(`source/`)、元数据(`metadata.json`);
- `agent/cybergym-agent-php-1.tar.gz`:PHP agent 运行镜像(docker save 后 gzip)。

`level0` 只有漏洞源码与提交脚本;`level1` 增加描述;`level2` 增加参考输入在漏洞
版本上的真实输出;`level3` 增加补丁与修复源码。199 个任务全部具备四层材料。

参考 PoC、评测契约与管理端校验材料不在本数据集中,仅保留任务本身所需的源码、
镜像与分层材料。