Skip to main content
BenchLM
Data
Specialist vertical · BenchAlign v6.0 · 47 ranked (13 Supported, 34 Estimated)
Scored from the dataset of

Cybersecurity leaderboard

As of November 21, 2026, Gemini 4 Carbon leads our cybersecurity leaderboard at 90.0 (Estimated, $2/$10 per 1M tokens), ahead of Claude Fable 5.2 at 87.0. GPT-5.5 is the highest Supported row at #8 (64.7); Qwen3.8 Max leads open-weight models at #10 (61.6). We rank 47 models (13 Supported, 34 Estimated) across 7 domain protocols, and this vertical carries zero weight in the overall score.

Shortlist by workload and budget

Each pick comes directly from the live cybersecurity rows below, with listed API rates and the case where another option wins.

Highest point score#1 · Estimated

Google · $2/$10 per 1M tokens

Carries an Estimated label; pick GPT-5.5 (#8, 64.7) when you require multi-source Supported coverage.

Highest Supported row#8 · Supported

OpenAI · $5/$30 per 1M tokens

Trails Gemini 4 Carbon by 25.3 points, so you trade peak point score for broader independent verification.

Top open-weight model#10 · Estimated

Alibaba · Open weights · self-hostable

Trails Gemini 4 Carbon by 28.4 points (#10 overall here), so pick it when weight control or on-prem residency outweighs peak score.

Ranked models

Ordered by BenchAlign v6.0 point estimate as of November 21, 2026. An Estimated row backed by fewer than two source families never ranks above a row scoring within 2 points below it. Conditional ranges do not establish rank confidence.

Cybersecurity LLM Leaderboard
RankModelScoreEvidence
#1Google90.0Estimated
#2Anthropic87.0Estimated
#3OpenAI82.0Estimated
#4Google79.3Estimated
#5Anthropic73.4Estimated
#6OpenAI70.3Estimated
#7Meta68.8Estimated
#8OpenAI64.7Supported
#9Thinking Machines Lab63.9Estimated
#10
Qwen3.8 MaxOpen weights
Alibaba
61.6Estimated
#11
Kimi K3Open weights
Moonshot AI
59.4Supported
#12OpenAI59.4Supported
#13
Kimi K3.1 ProOpen weights
Moonshot AI
61.3Estimated
#14
GLM-5.3Open weights
Z.AI
58.8Estimated
#15Google60.8Estimated
#16Google58.7Supported
#17Anthropic58.7Estimated
#18Google58.2Estimated
#19Anthropic56.2Supported
#20Anthropic56.2Supported
#21Alibaba56.2Estimated
#22OpenAI56.0Estimated
#23
GLM-5.2Open weights
Z.AI
55.4Estimated
#24Anthropic52.5Supported
#25
DeepSeek V4 ProOpen weights
DeepSeek
54.4Estimated
#26xAI53.9Estimated
#27OpenAI53.7Estimated
#28
GLM-5.1Open weights
Z.AI
50.7Supported
#29Meta50.5Estimated
#30Anthropic48.1Supported
#31Anthropic48.1Supported
#32Google49.9Estimated
#33OpenAI49.7Estimated
#34
InklingOpen weights
Thinking Machines Lab
48.8Estimated
#35xAI45.6Estimated
#36Anthropic44.3Estimated
#37Anthropic44.3Estimated
#38
Kimi K2.5Open weights
Moonshot AI
42.9Estimated
#39
GLM-4.7Open weights
Z.AI
41.4Estimated
#40Google41.3Estimated
#41
Beam 1.0 501B-A23BOpen weights
Reflection AI
40.2Estimated
#42
GLM-5Open weights
Z.AI
37.6Supported
#43xAI37.7Estimated
#44Meta37.0Estimated
#45Anthropic33.4Supported
#46Anthropic33.4Supported
#47
DeepSeek V3.2Open weights
DeepSeek
35.2Estimated

Scored domain benchmarks (7)

Each percentage is the protocol's relative reference weight inside this vertical's calibrated BenchAlign v6.0 model, not a raw score average. Independent evaluator shares are capped at 40% whenever three or more independent evaluators contribute.

A cybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.

32% ref. weight

A controlled benchmark for evaluating whether AI agents can extend vulnerability-triggering inputs into working exploits.

27% ref. weight
PromptArmor Frontier Breach· PromptArmor

Enterprise indirect prompt-injection defense across untrusted documents, emails, and tool outputs.

13% ref. weight

A cybersecurity benchmark for evaluating LLM agents on full-control V8 exploit synthesis using 16 measured exploit capability flags.

10% ref. weight

Quesma's BinaryAudit asks AI agents to find backdoors and timebombs planted in compiled, stripped binaries of real server software, using the reverse engineering tools Ghidra and radare2. The score here is the share of detection runs that found the planted code.

10% ref. weight

A cybersecurity benchmark for evaluating LLM agents on full-control V8 exploit synthesis using 16 measured exploit capability flags.

5% ref. weight

Measures how often indirect prompt-injection attacks succeed within fifteen attempts. Lower scores mean greater resistance to the tested attacks.

5% ref. weight

Patch reproduction, exploit synthesis, and prompt-injection defense sit on one calibrated scale

This board combines seven open and benchmark-owner security protocols led by CyberGym (32%), ExploitGym (27%), PromptArmor Frontier Breach (13%). CyberGym and BinaryAudit test whether an agent can locate and reproduce real software vulnerabilities inside containerized targets, ExploitGym and ExploitBench test whether it can turn crash inputs into working exploits, and PromptArmor Frontier Breach plus Gray Swan IPI measure whether untrusted inputs hijack the model.

We calibrate each protocol with BenchAlign v6.0 item-response difficulty parameters instead of averaging raw percentages, because a 30% solve rate on containerized exploit tasks is harder than an 85% defense rate on prompt-injection suites. Every scored row must also qualify for the general catalog, which keeps narrow single-task harnesses from distorting the scale.

Why we isolate cybersecurity from the overall score and exclude restricted cyber profiles

Dual-use offensive capability never moves our overall ranking: a model does not climb the main leaderboard by writing better V8 exploits. Vendor-gated cyber profiles such as GPT-6 Cyber under OpenAI Daybreak Red and Claude Mythos under Anthropic's verification program are also left off this general-access table because most engineering teams cannot call them over standard API keys.

A high score here does not clear a model for autonomous SOC triage, production firewall rules, or unreviewed code patching, and it never authorizes offensive testing. When you need per-benchmark numbers including restricted-access models, use the task tables on /cybersecurity rather than a composite score.

Choosing between frontier APIs and self-hosted security models

Gemini 4 Carbon (90.0, $2/$10 per 1M tokens) leads Claude Fable 5.2 (87.0, $8/$40 per 1M tokens) by 3.0 points, while GPT-5.5 (#8, 64.7) is the highest row with Supported multi-evaluator coverage. Because Gemini 4 Carbon already prices at $2/$10 per 1M tokens, you do not have to pay flagship $8/$40 token rates to get the top point estimate on this board.

Teams auditing proprietary repositories that cannot leave their own VPC should start with Qwen3.8 Max (#10, 61.6), then budget for the 28.4-point capability gap against Gemini 4 Carbon by pairing the model with deterministic static-analysis scanners and unit-test gates.

Questions

Which LLM is best for cybersecurity right now?

Gemini 4 Carbon holds #1 on our cybersecurity leaderboard at 90.0 (Estimated) as of November 21, 2026, ahead of Claude Fable 5.2 at 87.0 and GPT-6 Astra at 82.0. GPT-5.5 (#8, 64.7) is the highest Supported option when independent multi-evaluator depth matters most.

What is the best open-weight LLM for cybersecurity?

Qwen3.8 Max from Alibaba leads open-weight models on this cybersecurity board at #10 with a calibrated score of 61.6 (Estimated). That trails Gemini 4 Carbon by 28.4 points, so choose Qwen3.8 Max when self-hosting, VPC data residency, or weight customization matters more than peak frontier accuracy.

Which API model offers the best value for cybersecurity?

Gemini 4 Carbon leads both point accuracy (90.0) and price efficiency here at $2/$10 per 1M tokens, undercutting $8/$40 flagship tiers without sacrificing cybersecurity quality. If you need an even lower input rate for high-volume batch runs, compare mid-tier alternatives in the table below against your own retry rate.

Which benchmarks make up the cybersecurity leaderboard score?

BenchAlign v6.0 calibrates 7 open and primary domain protocols for this board, led by CyberGym (32%), ExploitGym (27%), PromptArmor Frontier Breach (13%), ExploitBench (official board) (10%). Each percentage reflects share of the vertical's reference weight inside our item-response model rather than a raw score average, and no single independent evaluator can exceed 40% when three or more evaluators contribute.

Why are GPT-6 Cyber and Claude Mythos absent from this cybersecurity leaderboard?

This leaderboard only ranks general-access models eligible for the main catalog, so engineering teams can deploy what they see. Vendor-gated cyber profiles such as GPT-6 Cyber (Daybreak Red) and Claude Mythos require program approval and run under custom lab harnesses; we track their per-benchmark results separately on /cybersecurity.

Do cybersecurity benchmark scores affect the overall LLM ranking?

No. Cybersecurity is an isolated specialist vertical with zero weight in our overall score, ensuring dual-use exploit generation never boosts a model's general rank. Among the 47 models here, 13 hold Supported labels and 34 hold Estimated labels, with single-source estimates held behind close Verified contenders.

Cybersecurity leaderboard updates

Weekly updates when new models or domain evaluations move the cybersecurity rankings.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.