File size: 4,824 Bytes
7124eb0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc46053
7124eb0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc46053
7124eb0
cc46053
7124eb0
 
 
 
 
 
 
 
 
cc46053
7124eb0
cc46053
7124eb0
cc46053
7124eb0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc46053
7124eb0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc46053
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
106
107
108
109
110
111
---
pretty_name: Independent ExploitBench Results
language:
  - en
tags:
  - exploitbench
  - v8-bench
  - cybersecurity
  - llm-agents
  - ai-agents
  - benchmark-results
  - agent-evaluation
  - software-exploitation
  - vulnerability-research
  - chromium-v8
  - javascript-engine
  - webassembly
  - exploit-synthesis
  - capability-ladder
  - transcripts
  - tool-calls
---

# Independent ExploitBench Results

This dataset contains **independent ExploitBench v8-bench evaluation results** for LLM cybersecurity agents. It makes model-level results, capability-ladder scores, run metadata, transcripts, and tool-call traces easy to find, compare, audit, and reproduce.

**This is an unofficial, independent results repository.** It is not maintained by the ExploitBench authors, Carnegie Mellon University, or the official ExploitBench organization.

## About ExploitBench

[ExploitBench](https://exploitbench.ai/) is a capability-ladder benchmark for evaluating how far LLM cybersecurity agents progress through real software-exploitation tasks: from reaching vulnerable code and reproducing a crash to building exploit primitives and achieving arbitrary code execution (ACE).

The first benchmark instance, **v8-bench**, evaluates real N-day vulnerabilities in Chromium's V8 JavaScript and WebAssembly engine. The benchmark defines 16 deterministically graded exploitation capabilities across five tiers, without relying on an LLM judge.

Canonical resources:

- [Official ExploitBench website](https://exploitbench.ai/)
- [ExploitBench paper on arXiv](https://arxiv.org/abs/2605.14153)
- [Official ExploitBench code](https://github.com/exploitbench/exploitbench)
- [Official ExploitBench Hugging Face dataset](https://huggingface.co/datasets/exploitbench/v8)

## Dataset contents

The dataset includes:

- aggregate results by model, provider, run, vulnerability, and capability tier;
- per-environment capability grades and the highest capability reached;
- complete agent transcripts and structured tool-call logs where releasable;
- benchmark, harness, container, and environment revisions;
- model identifiers, inference settings, budgets, seeds, and run dates;
- token usage, cost, timing, and failure information when available;
- reproduction and audit status for each result.

Release-specific schema and provenance metadata document the exact columns, splits, and evaluation coverage.

## Evaluation metadata

Each result includes enough context to make comparisons meaningful:

| Field | Description |
|---|---|
| Model | Provider, display name, and served model ID |
| Benchmark revision | ExploitBench commit, tag, or release |
| Environment | v8-bench target/CVE, image digest, and build revision |
| Harness | Agent harness and version |
| Budget | Turn, token, time, and cost limits |
| Assistance | Coaching, hints, or AutoNudge configuration |
| Repetition | Seeds and number of attempts |
| Outcome | Capability bitmap, highest tier, score, and failure reason |
| Provenance | Run date, artifact hashes, and audit status |

## Loading the dataset

Load the dataset with the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("shirman/exploitbench-results")
print(dataset)
```

## Responsible use and benchmark integrity

ExploitBench concerns real vulnerability exploitation. Use these materials only for authorized security research, defensive evaluation, reproducibility, and model-safety work. Do not use them to compromise systems or data you do not own or have explicit permission to test.

To reduce benchmark contamination, do not train or fine-tune models on held-out benchmark targets or result traces and then present those models as independently evaluated on the same targets. Disclose any prior exposure, training use, or prompt leakage.

## Keywords

ExploitBench, v8-bench, LLM cybersecurity agents, AI agent evaluation, Chromium V8, JavaScript engine security, WebAssembly security, N-day vulnerabilities, CVE exploitation, software exploitation, exploit synthesis, capability ladder, deterministic grading, exploit primitives, arbitrary read/write, control-flow hijack, sandbox escape, arbitrary code execution, ACE, benchmark transcripts, and tool-call traces.

## Citation

Please cite the original ExploitBench paper when using the benchmark:

```bibtex
@misc{lee2026exploitbench,
  title         = {ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents},
  author        = {Seunghyun Lee and David Brumley},
  year          = {2026},
  eprint        = {2605.14153},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  url           = {https://arxiv.org/abs/2605.14153}
}
```

When citing these results, include the repository URL and an immutable Hugging Face revision alongside the benchmark citation.