Dreadnode

// Research · DreadIndex

Benchmarking language models against offensive cyber security evaluations.

For each task, independent analyzers.Explore how we score and analyze model performance.

For more information on methodology and evaluation criteria, check out our writeup or contact the Dreadnode research team.

Max cost Any cost

Leaderboard

0 models · Release 11 · Updated October 2026

#
Loading leaderboard data...

Score vs. Cost

DreadIndex score versus total run cost
ModelScoreRun costOn value frontier

Score by Release Date

DreadIndex score by model release date
ModelScoreReleasedOn capability frontier

Hover a point for a quick summary. Click or tap it for the full breakdown.

Cite DreadIndex

Data release 11Published October 30, 2026

Source: DreadIndex 11 by Dreadnode (October 2026), https://dreadnode.io/research/dreadindex/

Evaluation analysis pipeline

Interactive pipeline

Select any node to inspect its details →

Run-scoped orchestration

Independent Analysis Runners

Reference

Shared Execution Controls and Result Assembly

Default LLM-Judge limits

Cheating
80
Scope
80
Refusal
20
Evaluation error
20

Maximum active 200

Default judge request policy

Model
glm-5.2
Timeout
120 sec
Temporary failures
3 attempts
Malformed output
2 corrections

Independent result collection [all five analyzers]

Independent
Analyzers never infer from or override each other
Full fidelity
Raw + parsed attempts and judge views retained
Explicit states
Incomplete states remain explicit, never negative
Durable output
Atomic writes preserve work through interruption

Run-scoped orchestration

Analysis Run Controller

  • validates run identity + configuration
  • owns lifecycle, progress + cancellation

Cite evaluation analysis pipeline

Source: DreadIndex evaluation analysis pipeline by Dreadnode, https://dreadnode.io/research/dreadindex/#pipeline

About DreadIndex

DreadIndex is a comprehensive benchmark designed to evaluate the offensive cyber security capabilities of language models. Developed by Dreadnode, it measures how well models perform across a range of adversarial security tasks including vulnerability discovery, exploit generation, and red team operations.

Each model is evaluated against a standardized set of challenges that reflect real-world offensive security scenarios. Scores represent overall performance across all evaluation categories, normalized to a 0-100 scale.

Each task also passes through the evaluation analysis pipeline: independent analyzers that look for cheating, scope violations, refusals, and evaluation errors, and record performance.