Last verified 2026-11-08

CybersecurityVendor-reportedPrices verified

ExploitGym leaderboard

As of November 8, 2026, Gemini 4 Carbon (Google DeepMind) leads ExploitGym at 54.8%, ahead of GPT-6 Astra at 42.4% and GPT-5.6 Sol at 30.3%. 9 systems ranked. Every score here is published by the model's own lab; each row links to the page it comes from.

See the ranking →All benchmarks →Model prices →

Top score

54.8%

Gemini 4 Carbon

Systems ranked

one setting

Independent rows

0/9

all vendor-reported

Cheapest near top

$1.21/1M

Gemini 4 Carbon · within 5 pts

With API prices

9/9

rates from our price index

Newest score

Nov 10

2026 · newest result

What is ExploitGym? ExploitGym measures whether an AI agent can turn a known software vulnerability into a working exploit, across userspace programs, Google's V8 engine and the Linux kernel. Made by UC Berkeley (Berkeley RDI) with academic and industry co-authors, May 2026.

§ 01 / LEADERBOARD

Top models on ExploitGym, November 2026

9 entries

#	Model	Score	Price / 1M	Context	Source
1	Gemini 4 Carbon Google DeepMind	54.8% via the Fairwind Program	$2 / $10	1M	Vendor-reported Google DeepMind 2026-11-10
2	GPT-6 Astra OpenAI	42.4% without production safeguards	$10 / $50	1.05M	Vendor-reported OpenAI 2026-09-03
3	GPT-5.6 Sol OpenAI	30.3% without production safeguards	$4 / $20	1.05M	Vendor-reported OpenAI 2026-09-03
4	Kimi K3.1 Pro Moonshot (Kimi)	20.7% max effort	$2 / $10	1M	Vendor-reported Moonshot AI 2026-11-02
5	MiMo-V2.6-Pro Xiaomi MiMo	17.8% as reported	$0.435 / $0.87	1M	Vendor-reported Xiaomi 2026-09-22
6	DeepSeek V4.1 Flash DeepSeek	15.3% pass@1	$0.3 / $1.2	1M	Vendor-reported DeepSeek 2026-09-10
7	MiMo-V2.6-Flash Xiaomi MiMo	6% as reported	$0.14 / $0.28	1M	Vendor-reported Xiaomi 2026-09-22
8	DeepSeek V4 Pro DeepSeek	5.4% pass@1	$1.32 / $3.96	1M	Vendor-reported DeepSeek 2026-09-10
9	DeepSeek V4 Flash Deprecated DeepSeek	1.8% pass@1	$0.44 / $1.32	1M	Vendor-reported DeepSeek 2026-09-10

§ 02 / SCORE VS PRICE

ExploitGym: score against price

The dashed line joins the models that no cheaper model beats: MiMo-V2.6-Flash (6%, $0.0477 per 1M tokens), MiMo-V2.6-Pro (17.8%, $0.144 per 1M tokens) and Gemini 4 Carbon (54.8%, $1.21 per 1M tokens). Anything below the line costs more for a lower score. Price is the effective rate per 1M tokens on current list prices: 92% input, 8% output, 82% of input served from cache.

§ 03 / PROVENANCE

Can you trust ExploitGym scores?

No independent leaderboard publishes ExploitGym results for these models yet, so every number on this page is the lab's own report of its own model. Labs choose the harness, the effort level and the subset they report, and they rarely publish failures. Read the setting under each score before comparing two rows.

!Each instance can run with or without standard security protections, and the authors show protections change success rates. A score means little without the configuration.
!The task is dual-use: the authors frame strong scores as a growing cybersecurity risk, not only a capability.
!There is no official public leaderboard; every score here is a lab's own report.

§ 04 / HOW IT WORKS

How ExploitGym works

The task

The agent gets vulnerable source code, an input that triggers the bug and a containerised runtime, and has to extend that input into an exploit that reaches unauthorised file access or code execution.

The score

An instance is solved when the exploit captures a secret flag that legitimate use cannot reach. Scores are the share of instances solved.

Test set

898 instances from real vulnerabilities, as published by the authors.

Made by

UC Berkeley (Berkeley RDI) with academic and industry co-authors, first published May 2026.

Official sources

§ 05 / SOURCES

Sources and freshness

9 published scores from 5 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.

Vendor-reported DeepSeek 3 scores · newest 2026-09-10
Vendor-reported OpenAI 2 scores · newest 2026-09-03
Vendor-reported Xiaomi 2 scores · newest 2026-09-22
Vendor-reported Google DeepMind 1 score · newest 2026-11-10
Vendor-reported Moonshot AI 1 score · newest 2026-11-02

ExploitGym FAQ

Answers derived from the 9 published scores on this page and the current price index.

Q · 01

Which AI model has the highest ExploitGym score?

Gemini 4 Carbon with 54.8% (via the Fairwind Program), as reported by Google DeepMind on November 10, 2026. Next: GPT-6 Astra at 42.4% and GPT-5.6 Sol at 30.3%.

Q · 02

What is the cheapest model near the top of ExploitGym?

The leader itself. No priced model within 5 points of Gemini 4 Carbon is cheaper: it runs at $1.21 per 1M tokens on our effective blend ($2 input, $10 output).

Q · 03

Are these ExploitGym scores independently verified?

Not yet. Every score here is the model maker's own report. Each row names the lab and links to the page where it published the number.

Q · 04

What does ExploitGym measure?

ExploitGym measures whether an AI agent can turn a known software vulnerability into a working exploit, across userspace programs, Google's V8 engine and the Linux kernel. The agent gets vulnerable source code, an input that triggers the bug and a containerised runtime, and has to extend that input into an exploit that reaches unauthorised file access or code execution.

Q · 05

How is the ExploitGym score calculated?

An instance is solved when the exploit captures a secret flag that legitimate use cannot reach. Scores are the share of instances solved. The test set has 898 instances from real vulnerabilities.

Q · 06

How current is this ExploitGym leaderboard?

The newest score was published on November 10, 2026. Every row shows the date its source published it, and we last read all of them on November 8, 2026. Prices come from our model price index, where each rate carries the date it was read on the vendor's page.