WhitzardNUWA Lab ↗
NUWANUWA INDEX

Search NUWA Index

Research · Benchmarks · Findings · Perspectives

中文
← NUWA Index

Source record

AgentCyberRange

Published evaluation results collected from AgentCyberRange, retaining the reported systems and configurations.

Original source and provenance ↓

Evaluation results

Reported by: AgentCyberRange website · Executed by: Not established for every submission · Captured: 2026-10-26

Web exploitation and post-exploitation are separate evaluations. Pass@3 retains the site’s definition; row-specific budgets, sample counts and execution dates are not disclosed.

11 of 11 systems shown in source order

Success rate (Pass@3) · Web exploitation · level0

AgentCyberRangeSuccess rate (Pass@3) · Web exploitation · level0Source order · native units · configuration-specific0255075100GPT-5.6 Sol / Codex 0.144.1: 35.37%; Web exploitation; Budget not specifiedGPT-5.6 Sol / Codex 0.144.135.37%GPT-5.5 / Codex 0.133.0: 18.97%; Web exploitation; Budget not specifiedGPT-5.5 / Codex 0.133.018.97%GPT-5.5 / Codex 0.147.0: 26.67%; Web exploitation; Budget not specifiedGPT-5.5 / Codex 0.147.026.67%Claude-Opus-4.7 / Claude Code 2.1.150: 14.02%; Web exploitation; Budget not specifiedClaude-Opus-4.7 / Claude Code 2.1.15014.02%GLM-5.1 / Claude Code 2.1.150: 11.97%; Web exploitation; Budget not specifiedGLM-5.1 / Claude Code 2.1.15011.97%DeepSeek-V4-Pro / Claude Code 2.1.150: 10.26%; Web exploitation; Budget not specifiedDeepSeek-V4-Pro / Claude Code 2.1.15010.26%DeepSeek-V4-Flash / Claude Code 2.1.226: 14.19%; Web exploitation; Budget not specifiedDeepSeek-V4-Flash / Claude Code2.1.22614.19%Qwen-3.7-Max / Qwen Code 0.16.1: 13.51%; Web exploitation; Budget not specifiedQwen-3.7-Max / Qwen Code 0.16.113.51%Qwen-3.8-Max / Qwen Code 0.21.10: 22.39%; Web exploitation; Budget not specifiedQwen-3.8-Max / Qwen Code 0.21.1022.39%Kimi-2.6 / Kimi Code 1.44.0: 5.81%; Web exploitation; Budget not specifiedKimi-2.6 / Kimi Code 1.44.05.81%Kimi-K3 / Kimi Code 0.35.0: 24.61%; Web exploitation; Budget not specifiedKimi-K3 / Kimi Code 0.35.024.61%https://agentcyberrange.io/
  1. Codex · Budget not specified
  2. Codex · Budget not specified
  3. Codex · Budget not specified
  4. Claude Code · Budget not specified
  5. Claude Code · Budget not specified
  6. Claude Code · Budget not specified
  7. Claude Code · Budget not specified
  8. Qwen Code · Budget not specified
  9. Qwen Code · Budget not specified
  10. Kimi Code · Budget not specified
  11. Kimi Code · Budget not specified
Source-reported systems and evaluation conditions
SystemResultTask set / modeConfigurationDates / attribution
GPT-5.6 Sol / Codex 0.144.135.37%Web exploitationlevel0CodexNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
GPT-5.5 / Codex 0.133.018.97%Web exploitationlevel0CodexNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
GPT-5.5 / Codex 0.147.026.67%Web exploitationlevel0CodexNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Claude-Opus-4.7 / Claude Code 2.1.15014.02%Web exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
GLM-5.1 / Claude Code 2.1.15011.97%Web exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
DeepSeek-V4-Pro / Claude Code 2.1.15010.26%Web exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
DeepSeek-V4-Flash / Claude Code 2.1.22614.19%Web exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Qwen-3.7-Max / Qwen Code 0.16.113.51%Web exploitationlevel0Qwen CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Qwen-3.8-Max / Qwen Code 0.21.1022.39%Web exploitationlevel0Qwen CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Kimi-2.6 / Kimi Code 1.44.05.81%Web exploitationlevel0Kimi CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Kimi-K3 / Kimi Code 0.35.024.61%Web exploitationlevel0Kimi CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
GPT-5.6 Sol / Codex 0.144.174.1%Post exploitationlevel0CodexNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
GPT-5.5 / Codex 0.133.031.71%Post exploitationlevel0CodexNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
GPT-5.5 / Codex 0.147.042.07%Post exploitationlevel0CodexNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Claude-Opus-4.7 / Claude Code 2.1.15015.04%Post exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
GLM-5.1 / Claude Code 2.1.15011.37%Post exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
DeepSeek-V4-Pro / Claude Code 2.1.15012.2%Post exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
DeepSeek-V4-Flash / Claude Code 2.1.22628.46%Post exploitationlevel0Claude CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Qwen-3.7-Max / Qwen Code 0.16.113.02%Post exploitationlevel0Qwen CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Qwen-3.8-Max / Qwen Code 0.21.1027.64%Post exploitationlevel0Qwen CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Kimi-2.6 / Kimi Code 1.44.05.68%Post exploitationlevel0Kimi CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified
Kimi-K3 / Kimi Code 0.35.043.9%Post exploitationlevel0Kimi CodeNot specified · ? trialsNot specifiedSource row date; semantics unspecified · AgentCyberRange website · Executor: Not specified

Scope and attribution

Reported by
AgentCyberRange website
Actual evaluator
Not established for every observation
Captured version
2026-10-26
Source identity
5344f595968f9c3e

Web exploitation and post-exploitation are separate evaluations. Pass@3 retains the site’s definition; row-specific budgets, sample counts and execution dates are not disclosed.

Source findings

The following statements summarize the source. They are not independent NUWA experimental findings.

Official website hero identifies GPT-5.6 Sol with Codex as first on both tracks; leaderboard backing tokens are web 35.37 and post 74.1, displayed as 35.37% and 74.10%.

routes-B-0Mubrq.js::az.en.hero.result; QR/0/level0/pass3; $R/0/level0/pass3

Source-reported leaderboard ranking; not independently reproduced.

Overall-results caption scopes its curves to an original six-system evaluation: solid Pass@3 Avg, dashed Pass@3 Max for top two systems, and best-to-worst shading over three independent runs at each step budget; caption endpoints for GPT-5.5/Codex are 16.1% web and 31.7% post.

routes-B-0Mubrq.js::az.en.sections.overall.caption

Caption applies to original static figure, not all current leaderboard rows. Numeric curve series unavailable; no digitization.

Official page describes 110 vulnerabilities in 15 web applications and 156 internal hosts in eight enterprise-like ranges; these are benchmark-scope counts.

routes-B-0Mubrq.js::az.en.hero.body

Scope counts do not establish per-row sample denominators.

Extraction coverage

Each entry records the treatment of the source section. A reviewed observation does not imply that every claim or result in the document was verified.

View 18 coverage entries
SectionTreatmentLocation and scope
Official homepage HTML shellextractedhomepage.htmlMetadata and modulepreload identify routes-B-0Mubrq.js; the static shell does not contain table cells.
Web leaderboard backing recordsextractedroutes-B-0Mubrq.js::QR11 source-order rows, all 70 numeric fields, including 17 top-level Level-0 aliases; no missing values imputed.
Post leaderboard backing recordsextractedroutes-B-0Mubrq.js::$R13 source-order rows, all 82 numeric fields. Two agent entries (Strix, Pentagi) retained but quarantined because harness version is not disclosed.
Model-keyed version map and per-row overrideextractedroutes-B-0Mubrq.js::ez10 map entries. Explicit GPT-5.5 row agentVersion=0.147.0 overrides model map 0.133.0. No model snapshot/API version disclosed.
Identity, union, formatting, sorting and metadataextractedroutes-B-0Mubrq.js::function hzLevel-0 Pass@3 rendered in both columns; initial sort is Post descending. Submitter and submitted dates are renderer-derived, not evaluation dates.
Step budget summaryqualitativeroutes-B-0Mubrq.js::function hz/summary metricsExact literal summary 150 / 500 retained. Mapping to tracks or specific rows is not established by this renderer; budgets remain null.
Web exploitation success by levelreusedroutes-B-0Mubrq.js::cz18 exact values across Level-0/1/2 and original six models; each verified equal to its QR nested pass3 counterpart.
Post exploitation success by levelreusedroutes-B-0Mubrq.js::sz18 exact values across Level-0/1/2 and original six models; each verified equal to its $R nested pass3 counterpart.
Success rate over execution stepsfigure-onlyroutes-B-0Mubrq.js::az.en.sections.overallOfficial static asset /assets/overall-result-aVhp3QZG.png. Caption describes original six-system evaluation, Pass@3 Avg/Max and three independent runs. Curve datapoints not numerically extracted or estimated.
Overall-figure caption endpointsqualitativeroutes-B-0Mubrq.js::az.en.sections.overall.captionCaption reports GPT-5.5/Codex 16.1% web and 31.7% post for original evaluation. Different context from current leaderboard; retained as claim and conflict, not normalized into leaderboard.
Benchmark scopequalitativeroutes-B-0Mubrq.js::az.en.hero.body / iz110 vulnerabilities/15 web applications; 156 internal hosts/eight ranges; 16 vulnerability classes described in tracks.web.detail. Dataset sizes do not establish row denominators.
Unrendered cost/time fieldsextractedroutes-B-0Mubrq.js::QR; routes-B-0Mubrq.js::$R28 values retained and quarantined: current page gives no units or computation definitions.
Results analysisqualitativeroutes-B-0Mubrq.js::az.en.findingsHigh-level exploration/information-gathering limitation only in derived package prose. Original website source capture retained for provenance; raw task datasets, payloads and credentials are not collected.
Level-0/1/2 protocol definitionsnot-disclosedroutes-B-0Mubrq.js::QR; routes-B-0Mubrq.js::$R; az.en.sections.difficultyLevel labels available, but their precise assistance/condition definitions are not stated in the extracted page; do not infer difficulty ordering or assistance.
Row-specific trials/sample counts/evaluation datesnot-disclosedroutes-B-0Mubrq.js::QR; routes-B-0Mubrq.js::$RNo per-row trials, success numerator/denominator, evaluation date, exact model snapshot, shared-budget guarantee, or shared-sample guarantee. Figure-caption three-run claim is scoped to original figure only.
Linked paperpendinghttps://arxiv.org/abs/2606.14295Dependency linked by official page; paper not opened or extracted in this website data unit.
Architecture figurenot-relevantroutes-B-0Mubrq.js::az.en.sections.archNot needed to extract scoreboard values; no architecture/raw task reconstruction performed.
Citation and affiliationqualitativeroutes-B-0Mubrq.js::oz; az.en.hero.affiliationWebsite cites arXiv:2606.14295, year 2026, and Fudan University. No publication/revision day is inferred.

Unresolved source issues

部分智能体记录缺少可核实的版本信息，已排除；当前榜单只纳入显示的 Level-0 Pass@3。

原始结果图与当前榜单的范围不同，图注数值分别保留，不当作同一条件的重复测量。

Version fingerprints
document

https://agentcyberrange.io/

d1c16a642875aa69e04798a9825870c7e1f7eb589649d659769ba607aa9a4f3b
data

https://agentcyberrange.io/assets/routes-B-0Mubrq.js

a03398a3e2fd33f2c9ccf41d959bbfc5d026238cad938bccf58a49d3c9cc0484
renderer

https://agentcyberrange.io/assets/routes-B-0Mubrq.js

a03398a3e2fd33f2c9ccf41d959bbfc5d026238cad938bccf58a49d3c9cc0484
data

https://agentcyberrange.io/assets/routes-B-0Mubrq.js

04a099d99579b876aa09688f0881a0c412aa9328295e2309f1f578c822ed91d4
renderer

https://agentcyberrange.io/assets/routes-B-0Mubrq.js

36c61bdc2a14e2201c1a36d79373d2f607879abab96fea84f2a352e00a045e6a
document

https://agentcyberrange.io/

3211a260a4489fdaa558a8503439713e58b03c368dddcc9870a0e537e2f7d1c1

References

Numerical results are reported by the cited authors.

  1. [1] AgentCyberRange: Evaluate Agent Cyberattack Capabilities in Enterprise-like Cyber Ranges