Agents · Benchmark profile

ExploitBench

A cybersecurity-agent benchmark measuring progress from reaching vulnerable code through exploit primitives and code execution.

Ranking weighted2Rolling3 models

Data verified 15 Jul 2026 · Methodology 2.3.0

Benchmark score on ExploitBench

Leaderboard · 3 models

GLM-5.3 leads at 54.4%, followed by GPT-5.5 (34%) and Claude Opus 4.7 (27%).

Ranking · 1.6%Agents2

#1Closed

GLM-5.3

Z.AI

glm-5.3

54.4%

Overall #3 · Context 1M tokens

#2Closed

GPT-5.5

OpenAI

gpt-5.5

34%

Overall #13 · Context 1.05M tokens

#3Closed

Claude Opus 4.7

Anthropic

claude-opus-4-7

27%

Overall #23 · Context 1M tokens

ViewModelsColours

Z.AIOpenAIAnthropic

Lumina

ExploitBench leaderboard

Accessible data for ExploitBench leaderboard
Model	Rank	Model ID	Provider	Score	Unit
GLM-5.3	#1	glm-5-3	Z.AI	54.4	percent
GPT-5.5	#2	gpt-5-5	OpenAI	34	percent
Claude Opus 4.7	#3	claude-opus-4-7	Anthropic	27	percent

Benchmark score table (3 models)

One best compatible score per canonical product · higher is better

ExploitBench leaderboard ranked by score
Rank	Model	Provider	License	Evidence use	Score
#1	GLM-5.3 glm-5.3	Z.AI	announced-open	Score input	54.4%
#2	GPT-5.5 gpt-5.5	OpenAI	closed	Score input	34%
#3	Claude Opus 4.7 claude-opus-4-7	Anthropic	closed	Score input	27%

The top of this snapshot is led by GLM-5.3 at 54.4%; third place is 27.4 points behind. The top-3 spread is 27.4 points.

Configuration-preserving evidencePublished source rows (9)Exact source labels, configurations, versions and dated observations. Same-snapshot canonical/configuration projections are shown once.

Every distinct dated source-published row for ExploitBench, including exact model configuration labels
Exact published model / configuration	Canonical product	Score	Version / system	Evidence	Source / observed
Claude Fable 5 w/ fallback as published by Z.AI claude-fable-5-max	Claude Fable 5	78%	2Z.AI GLM-5.3 official comparison table	Relative comparison · not standard direct provider-reported	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities ↗Observed 2026-08-14 · checked 2026-08-15
GPT-5.6 Sol as published by Z.AI gpt-5-6-sol-max	GPT-5.6 Sol	76.5%	2Z.AI GLM-5.3 official comparison table	Relative comparison · not standard direct provider-reported	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities ↗Observed 2026-08-14 · checked 2026-08-15
GLM-5.3 (max) glm-5-3-max	GLM-5.3	54.4%	2Z.AI GLM-5.3 official comparison table	Score input provider-reported	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities ↗Observed 2026-08-14 · checked 2026-08-15
Claude Opus 4.8 as published by Z.AI claude-opus-4-8-max	Claude Opus 4.8	40%	2Z.AI GLM-5.3 official comparison table	Relative comparison · not standard direct provider-reported	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities ↗Observed 2026-08-14 · checked 2026-08-15
Kimi K3 as published by Z.AI kimi-k3-max	Kimi K3	32.2%	2Z.AI GLM-5.3 official comparison table	Relative comparison · not standard direct provider-reported	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities ↗Observed 2026-08-14 · checked 2026-08-15
Qwen3.8 Max as published by Z.AI qwen-3-8-max-xhigh	Qwen3.8 Max	28.8%	2Z.AI GLM-5.3 official comparison table	Relative comparison · not standard direct provider-reported	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities ↗Observed 2026-08-14 · checked 2026-08-15
GLM-5.2 as published by Z.AI glm-5-2-max	GLM-5.2	24.4%	2Z.AI GLM-5.3 official comparison table	Relative comparison · not standard direct provider-reported	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities ↗Observed 2026-08-14 · checked 2026-08-15
Exact model variant as published on the source leaderboard page. source label without registered configuration ID	GPT-5.5	34%	2Public leaderboard extraction	Score input source-checked	ExploitBench via BenchLM ↗Observed 2026-07-15 · checked 2026-07-15
Exact model variant as published on the source leaderboard page. source label without registered configuration ID	Claude Opus 4.7	27%	2Public leaderboard extraction	Score input source-checked	ExploitBench via BenchLM ↗Observed 2026-07-15 · checked 2026-07-15

Compare top models

GLM-5.3 vs GPT-5.5 GPT-5.5 vs Claude Opus 4.7

About ExploitBench

Definition and scoring

Organisation: ExploitBench authors
Category: Agents
Version: 2
Direction: higher is better
Ranking use: Weighted 1.6%
Contamination risk: Low

A cybersecurity-agent benchmark measuring progress from reaching vulnerable code through exploit primitives and code execution. Every genuine source row stays tied to its exact model label, configuration, benchmark version and evaluation system. The summary chart shows one best compatible score per canonical product; Score 2.0 admits only explicitly mapped, frozen protocols and keeps incompatible configurations separate.

Open benchmark source ↗

FAQ

Quick answers

What does ExploitBench measure?

A cybersecurity-agent benchmark measuring progress from reaching vulnerable code through exploit primitives and code execution.

Which model scores highest on ExploitBench?

GLM-5.3 by Z.AI currently leads with 54.4%.

How many models are evaluated on ExploitBench?

3 canonical model products have a qualifying score, backed by 9 distinct dated source rows.

Does this affect overall Lumina rank?

Yes. This family is ranking-weighted at 1.6% of its capability category. GLM-5.3 is overall #3.

Agents · Benchmark profile

ExploitBench

A cybersecurity-agent benchmark measuring progress from reaching vulnerable code through exploit primitives and code execution.

Ranking weighted2Rolling3 models

Data verified 15 Jul 2026 · Methodology 2.3.0

Benchmark score on ExploitBench