ExploitGym leaderboard

As of November 14, 2026, among the results listed here, Google Gemini 4 Carbon has the highest reported ExploitGym score, 54.8% (10h), published by Google on October 28, 2026, ahead of Google Gemini 4 Argon at 42.0%. 10 models have a ExploitGym result in our records.

Scores as the model makers published them, not re-run by us; each row links to its announcement.

Best reported score

54.8%

Google Gemini 4 Carbon

Reported by Google, October 28, 2026 · 10h

Runner-up

42.0%

Google Gemini 4 Argon

Reported by Google, October 28, 2026 · 10h

10 models ranked · 4 makers

Reported ExploitGym scores

ExploitGym scores reported by model makers
Rank	Model	Score	Setting	Reported by	Date
1	Gemini 4 Carbon	54.8%	10h	Google	October 28, 2026 · read from the maker's chart
2	Gemini 4 Argon	42.0%	10h	Google	October 28, 2026 · read from the maker's chart
3	GPT-5.6 Sol	33.7%	Not published	OpenAI	August 7, 2025 (model release)
4	GPT-5.6 Terra	23.2%	Not published	OpenAI	August 7, 2025 (model release)
5	MiMo-V2.6-Pro	17.8%	Not published	Xiaomi	September 21, 2026 (model release)
6	DeepSeek V4.1 Flash	15.3%	Pass@1	DeepSeek	September 10, 2026 (model release)
7	GPT-5.5	15.1%	Not published	OpenAI	August 7, 2025 (model release)
8	GPT-5.6 Luna	12.4%	Not published	OpenAI	August 7, 2025 (model release)
9	MiMo-V2.6-Flash	6.0%	Not published	Xiaomi	September 21, 2026 (model release)
10	MiMo-V2.5-Pro	0.2%	Not published	Xiaomi	September 21, 2026 (model release)

Other reported runs

ExploitGym scores reported by model makers, other reported runs
Model	Score	Setting	Reported by	Date
Gemini 4 Argon	42.0%	Not published	Google	October 28, 2026 · read from the maker's chart
Gemini 4 Argon	39.5%	6h	Google	October 28, 2026 · read from the maker's chart
Gemini 4 Carbon	54.8%	Not published	Google	October 28, 2026 · read from the maker's chart
Gemini 4 Carbon	53.5%	6h	Google	October 28, 2026 · read from the maker's chart

How much to trust these numbers

Source: Reported by the model makers themselves. Not re-run by LLMBenchmark or anyone independent.
Settings: Settings differ between makers. Check the Setting column before comparing two scores.
Spread: 54.8% to 0.2% across 10 models.
Freshness: Newest result published October 28, 2026; oldest August 7, 2025.
Charts: Some values were read from charts in the announcements. Two independent readings had to match exactly.

What ExploitGym measures

ExploitGym tests whether AI agents can develop exploits for real-world vulnerabilities; the provided results do not explain what its score means. Source: ExploitGym