Papan peringkat

Peringkat CyberGym Level 1: setelah Agent menerima deskripsi kerentanan dan basis kode yang belum ditambal, Agent harus menghasilkan PoC yang berfungsi untuk mereproduksi kerentanan target.

Jenis:
Percobaan:
#Agent / ModelTingkat keberhasilanTanggalSumber
1
Gcsa Agent Agentuses grok-4.5 + grok-4.6
91.3%
2026-08-26GCSA
2
Crystalline Agentuses Claude Opus 4.6
89.6%
2026-06-08Independent researcher
3
MDASH Agentuses GPT-5.4, Claude Opus 4.6, Claude Sonnet 4.6
88.5%
2026-05-12Microsoft
4
GPT-5.5-Cyber ModelOpenAI Agent
85.6%
2026-06-22OpenAI
5
Xuanwu Atuin AI Agentuses GLM-5.2
84.8%
2026-07-22Tencent Xuanwu Lab
6
Claude Mythos Preview ModelAnthropic Agent
83.1%
2026-04-07Anthropic
7
GPT-5.5 ModelOpenAI Agent
81.8%
2026-04-23OpenAI
8
GPT-5.4 ModelOpenAI Agent
79.0%
2026-04-23OpenAI
9
MopMonk Agent Agentuses MiniMax M3
73.1%
2026-06-28MopMonk AI
10
Whitzard Agentuses GLM-5.1
68.9%
2026-07-21Whitzard Team
11
GLM-5.1 ModelClaude Code
68.7%
2026-04-07Zhipu AI
12
Claude Opus 4.6 ModelAnthropic Agent
66.6%
2026-02-05Anthropic
13
GPT-5.4 ModelCodex CLI
66.3%
2026-04-07Zhipu AI
14
Claude Sonnet 4.6 ModelAnthropic Agent
65.2%
2026-02-17Anthropic
15
SageAgent Agentuses GPT-5
60.2%
2026-02-09OpenSage Team
16
Muse Spark 1.1 ModelMeta Agent
59.0%
2026-07-09Meta AI
17
DeepSeek-V4-Pro ModelClaude Code
57.7%
2026-07-13XDxAI
18
Claude Opus 4.5 ModelAnthropic Agent
50.6%
2026-02-05Anthropic
19
Muse Spark ModelMeta Agent
43.5%
2026-04-14Meta AI
20
GLM-5 ModelClaude Code
43.2%
2026-02-12Zhipu AI
21
Kimi K2.5 ModelKimi Agent
41.3%
2026-02-02Kimi
22
GPT-5 ModelOpenHands
39.4%
2025-12-05CyberGym Team
23
Gemini 3.1 Pro ModelGemini CLI
38.8%
2026-04-07Zhipu AI
24
Claude Sonnet 4.5 ModelAnthropic Agent
28.9%
2025-09-29Anthropic
25
Claude Opus 4.1 ModelAnthropic Agent
25.0%
2025-09-29Anthropic
26
GLM-4.7 ModelClaude Code
23.5%
2026-02-12Zhipu AI
27
Claude Sonnet 4 ModelAnthropic Agent
22.6%
2025-09-29Anthropic
28
Claude Sonnet 4 ModelOpenHands
17.9%
2025-05-23CyberGym Team
29
Claude Sonnet 3.7 ModelAnthropic Agent
14.5%
2025-09-29Anthropic
30
Claude Sonnet 3.7 ModelOpenHands
11.9%
2025-05-15CyberGym Team
31
GPT-4.1 ModelOpenHands
9.4%
2025-05-15CyberGym Team
32
GPT-4.1 ModelCybench
9.0%
2025-05-15CyberGym Team
33
GPT-4.1 ModelCodex CLI
7.4%
2025-05-15CyberGym Team
34
GPT-4.1 ModelENiGMA
7.2%
2025-05-15CyberGym Team
35
Gemini 2.5 Flash Preview ModelOpenHands
4.8%
2025-05-15CyberGym Team
36
DeepSeek-V3 ModelOpenHands
3.6%
2025-05-15CyberGym Team
37
o4-mini ModelOpenHands
2.5%
2025-05-15CyberGym Team
38
R2E-Gym-32B ModelOpenHands
2.0%
2025-05-15CyberGym Team
39
Qwen3-235B-A22B ModelOpenHands
1.9%
2025-05-15CyberGym Team
40
OpenHands-LM-32B ModelOpenHands
1.7%
2025-05-15CyberGym Team
41
SWE-Gym-32B ModelOpenHands
0.1%
2025-05-15CyberGym Team

Data disinkronkan dari CyberGym · Tanggal cuplikan 2026-08-26

Katalog reproduksi kerentanan

Kumpulan tugas nyata CyberGym n=100 — AI Agent mereproduksi 78 kerentanan tanpa campur tangan manusia, mencakup 42 proyek sumber terbuka. Klik «Mulai» untuk masuk ke indeks proyek, lalu pilih proyek untuk melihat hasil pemindaian.

78

Berhasil direproduksi

42

Proyek sumber terbuka

11

Jenis CWE

78%

Tingkat keberhasilan

cryptidsmith://vulns/

Kami menjaga keamanan perangkat lunak yang menjalankan dunia

Klik entri untuk memperluas detail. Penilaian didasarkan pada pengiriman PoC ke penilai ASAN CyberGym dan hanya mengakui exit_code. HOT = terselamatkan setelah peningkatan sistem.