GreyNoise AI Cyber Index
A benchmark of AI capability grounded in defensive security work.
Compare how well AI models write detections and investigate attacks through evaluations built from real attack traffic observed by GreyNoise. Explore performance by task, refusal rates, time, and cost per problem.
Real attack data. Three defensive tasks.
Find the AI model for your security work.
Compare the models that can handle your work within your time and budget.
Models
All models
The data behind the benchmark.
GreyNoise helps defenders detect and respond to attacks on the network edge. It observes those attacks directly through its Global Observation Grid, more than 5,000 sensors in over 100 countries, as attackers scan, exploit, and compromise systems across the internet.
The GreyNoise AI Cyber Index is built from that real attack data, not synthetic environments. Every problem starts from an attack GreyNoise observed, and answers are graded against ground truth from labeled data GreyNoise uses in production.
That means the scores reflect how models handle the attacks defenders actually face.
Methodology reportHow we test AI for security workThe evaluations, the graders, and how every answer is scored on proof.Read the report (PDF) →What do the scores mean?
Each score is the share of available points a model earned, averaged across problems. 100% means full marks on every problem.
Open any model in the table to see its score on each evaluation.
How is the Cyber Index score calculated?
It combines the three evaluation scores, weighted .
How is each evaluation graded?
How do refusals impact scoring?
A refusal is an attempt where the model declined to answer.
Penalizes refusals on: a refused attempt scores zero.
Penalizes refusals off: refused attempts are left out, and each model is scored on the attempts it answered.
Refusal rates are shown either way.
Why show two sets of scores?
A model that refuses a security request can’t help with it, so refusals affect how reliably you can use a model in production.
With the toggle on, the score reflects skill and reliability. With it off, it shows skill on the attempts the model answered. Comparing the two shows how much refusals cost each model.
What does “capable” mean?
How is speed measured?
Average wall-clock time per attempt, including tool use. It is not token generation speed.
How is cost measured?
Average cost per problem at each provider’s published prices. Models GreyNoise runs itself are priced at GPU rental rates.
Time and cost don’t change with the refusal setting.
How are the fastest capable model and best value chosen?
Fastest capable: the capable model with the shortest time per attempt.
Best value: the capable model with the lowest cost per problem. Once any capable model solves 70% of problems, best value is picked from those models only.
Most attractive (shaded area): capable models at or below the median time or cost of the models shown.
Ties are judged to 0.1 point and marked with =. Missing values show as n/a, never as zero.