Skip to main content
Evaluation rates the models your agents depend on. Discovery finds the model; evaluation tells you how it behaved under attack.

​
Reading the health score

Every evaluated model carries a Health score. You meet it first as a column in Assets → Models, showing your models side by side. The health score is a grade from 0 to 100, and it runs the opposite direction from severity: a high severity is bad news, a high health score is good news. The Models table carries both, because they answer different questions: how the model held up under attack, and how serious the worst finding against it is. Each score falls into a band:
Not every model is evaluated. When no evaluation is available, the health score is empty rather than zero. This is expected.

​
What the category scores tell you

Behind the overall Model health score are category scores for Jailbreaks, Misuse, Toxicity, Security, and Robustness. Each summarizes how the model performed across the tests in that category, scored 0 to 100 the same way. A higher score is stronger performance under evaluation, not a guarantee of how the model behaves in your application. For what each category tests, see How evaluation works. Open an evaluated model to reach its Overview tab, where a Model health score card shows the overall score with the category scores beside it. The card appears only once a model has been evaluated, and a category with no result is left out rather than shown as zero.
A strong overall score can still conceal a weakness in one category. A model scoring 65 overall carries the Good band, but with Jailbreaks at 30 it is not a good choice for an agent that takes untrusted input — it is a jailbreak risk that the overall score doesn’t show you.

​
Probe results

Open a model and select its Evaluations tab to see the individual tests behind the score. Each row is one probe — a named attack technique — showing:
  • The probe name, and its category as a chip
  • What it tries, in a sentence
  • The score it earned, with its band
This is where a weak category score becomes something you can act on. A low Jailbreaks score is a list of named techniques the model did not hold up against, and the probe descriptions tell you what those techniques actually do.

​
What to do about a weak model

A low score is not automatically a problem. It is a problem in context.
1

Check which agents use it

Open the model and read Used by on its Overview tab, or open the Locations tab for the files each dependency sits in. A weak model nothing calls yet is a decision you get to make before it matters.
2

Match the weakness to the exposure

A low Toxicity score matters most where output reaches users. A low Jailbreaks score matters most where input comes from outside your organization. A weak score in a category your agent never exercises is a lower priority than a middling one it leans on.
3

Check whether it filed a finding

Scores past a threshold appear in Insights → Issues with a severity attached. If it’s there, it has already been triaged into your work queue — see Discovery results.
4

Swap it or contain it

Either move the agent to a stronger model, or leave it and constrain what reaches it with guardrails. If you’re swapping, read the Details card on the model’s Overview tab — Deprecation type, Suggested replacement, and Deprecation date tell you whether the model is on its way out and what replaces it.

​
Next steps

How evaluation works

What gets tested, how scoring works, and how a score becomes a finding

Discovery results

Triage findings, confirm they’re real, and know when they’re resolved