WebExploitBench
The benchmark’s paper hasn’t scored a model released after Qwen3.7-Max (May 2026). Labs report their own numbers for one newer model: Gemini 4 Neon at 51.3%.
Agents find and exploit 110 vulnerabilities in 15 real web applications, each in its own container. A verifier checks that the agent's proof of concept hits the same endpoint as the reference exploit.
The vulnerabilities are 18 zero-day, 56 one-day and 36 synthetic, across 17 kinds such as SQL injection and SSRF. The repo holds 5 of the 15 apps, and the full set is on Hugging Face behind a gate, for evaluation only. Google reports version 1.2.
- Made by
- AgentCyberRange
- Released
- June 2026
- Kind
- Finding and exploiting real bugs
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Risk
- Cyber offence
- Language
- English
- Interaction
- Agentic
- Items from
- Real work and use, Generated
- Refresh
- New versions replace it
- Run by
- Anyone
- In views
- Agents, Dangerous capabilities
- Strongest model tested
- GPT-5.5 #21 on the capability index
- Last in a launch post
- November 2026
- Last model added
- November 2026
- Scores refreshed
- November 20th, 2026
GPT-5.5 leads with 36.4% among models released by May 2026
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
Gemini 4 Neon leads with 51.3%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model. Striped bars are claims, and dotted bars are reruns by Epoch AI or another board.
The newest model with a score here came out in May 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 GPT-5.5 Level 1, xhigh 36.4%
- 2 Claude Opus 4.7 Level 2, max 29.1%
- 3
Qwen3.7-Max Level 1 26.4%
- 4 GLM-5.1 Level 2 17.3%
- 5
DeepSeek-V4-Pro Level 2 13.6%
- 6 Kimi K2.6 Level 1 12.7%
Source: The WebExploitBench paper
- 1
Gemini 4 Neon 51.3%
- 2 GPT-5.5 Level 1, xhigh 36.4%
- 3 Claude Opus 4.7 Level 2, max 29.1%
- 4
Qwen3.7-Max Level 1 26.4%
- 5 GLM-5.1 Level 2 17.3%
- 6
DeepSeek-V4-Pro Level 2 13.6%
- 7 Kimi K2.6 Level 1 12.7%
Sources: Labs’ model cards and the WebExploitBench paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 36.4% | Level 1, xhigh, Codex 0.133.0 | Apr 2026 | Official leaderboard |
| 2 | Claude Opus 4.7 | 29.1% | Level 2, max, Claude Code 2.1.150 | Apr 2026 | Official leaderboard |
| 3 | | 26.4% | Level 1, Qwen Code 0.16.1 | May 2026 | Official leaderboard |
| 4 | GLM-5.1 | 17.3% | Level 2, Claude Code 2.1.150 | Apr 2026 | Official leaderboard |
| 5 | | 13.6% | Level 2, Claude Code 2.1.150 | Apr 2026 | Official leaderboard |
| 6 | Kimi K2.6 | 12.7% | Level 1, Kimi Code 1.44.0 | Apr 2026 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | 51.3% | Nov 2026 | Lab model card | ||
| 2 | GPT-5.5 | 36.4% | Level 1, xhigh, Codex 0.133.0 | Apr 2026 | Official leaderboard |
| 3 | Claude Opus 4.7 | 29.1% | Level 2, max, Claude Code 2.1.150 | Apr 2026 | Official leaderboard |
| 4 | 26.4% | Level 1, Qwen Code 0.16.1 | May 2026 | Official leaderboard | |
| 5 | GLM-5.1 | 17.3% | Level 2, Claude Code 2.1.150 | Apr 2026 | Official leaderboard |
| 6 | 13.6% | Level 2, Claude Code 2.1.150 | Apr 2026 | Official leaderboard | |
| 7 | Kimi K2.6 | 12.7% | Level 1, Kimi Code 1.44.0 | Apr 2026 | Official leaderboard |
Labs report scores for 1 model in their cards. Switch the scores above to rank them together.
More effort adds a median 17 points
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The WebExploitBench paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.5 | Level 1, xhigh | Codex 0.133.0 | 36.4% | ||
| GPT-5.5 | Level 2, xhigh | Codex 0.133.0 | 31.8% | ||
| Claude Opus 4.7 | Level 2, max | Claude Code 2.1.150 | 29.1% | ||
| Level 1 | Qwen Code 0.16.1 | 26.4% | |||
| Claude Opus 4.7 | Level 1, max | Claude Code 2.1.150 | 24.6% | ||
| GPT-5.5 | Level 0, xhigh | Codex 0.133.0 | 19.1% | ||
| GLM-5.1 | Level 2 | Claude Code 2.1.150 | 17.3% | ||
| Claude Opus 4.7 | Level 0, max | Claude Code 2.1.150 | 16.4% | ||
| Level 2 | Claude Code 2.1.150 | 13.6% | |||
| GLM-5.1 | Level 1 | Claude Code 2.1.150 | 12.7% | ||
| Kimi K2.6 | Level 1 | Kimi Code 1.44.0 | 12.7% | ||
| Level 1 | Claude Code 2.1.150 | 12.7% | |||
| Level 2 | Qwen Code 0.16.1 | 12.7% | |||
| GLM-5.1 | Level 0 | Claude Code 2.1.150 | 11.8% | ||
| Level 0 | Qwen Code 0.16.1 | 10.9% | |||
| Level 0 | Claude Code 2.1.150 | 10% | |||
| Kimi K2.6 | Level 2 | Kimi Code 1.44.0 | 9.09% | ||
| Kimi K2.6 | Level 0 | Kimi Code 1.44.0 | 3.64% |
Number one over time 2 models have held number one with claims included since its release in June 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The WebExploitBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Official WebExploitBench leaderboard and labs’ model cards
One lab put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Nov 18, 2026 | 51.3 | WebExploitBench v1.2 (51.26% in the text, 51.3% in the table) |
One system card reports it
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Nov 18, 2026 | | 51.3 | WebExploitBench v1.2 (51.26% in the text, 51.3% in the table) (results table, p.4) | Cyber risk |
More in Cybersecurity
- ExploitGym In use 869 real-vulnerability exploit-development tasks (userspace, V8, Linux kernel). Success means capturing a flag through the intended vulnerability. GPT-6 Astra 42.4%
- ExploitBench In use 41 recent V8 JavaScript-engine N-day vulnerabilities. Progress measured by 16 grader-verified capability flags across five tiers. GPT-6 Astra 100%
- ExploitBench - Internal Port In use Agents develop working exploits for recently disclosed vulnerabilities to limit contamination. Scored as arbitrary code execution success rate. GPT-6.2 Astra 53.5%
- Atomic Challenges In use Agents solve 22 skill-level offensive security challenges in vulnerability research, network attack simulation and evasion. Scored as challenges solved. GPT-6 Astra 90.9%
- MiMo Cyber Bench In use Agents do in-house cybersecurity tasks that add to public vulnerability reproduction and exploitation benchmarks. The report gives no task count or grading details. MiMo-V2.6 Pro 81.7%
- PatchBench (OpenAI) In use Models patch security flaws. Automated tests and model-based review check that each patch fixes the flaw, keeps the intended behavior and is ready to merge. GPT-6.2 Astra 79.7%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Confidence check. Likely. Some facts rest on secondary sources. Every benchmark
Catalog record 1 warning 1 pass
- Pass Task access All tasks are public.
- Warning Confidence Likely. Some facts rest on secondary sources.
Scores 4 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (6 entries), launch posts (1 score) and system cards (1 score).
- Pass Models scored 7 models we track.
- Pass Newest score The newest is from November 2026, for Gemini 4 Neon.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
- Pass Headroom The best is 51.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 18 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.