Skip to main content
Catching a prompt injection vulnerability or a jailbreak path in production is expensive — it means a live agent is already exposed. Integrating Trident’s red-team evaluations into your CI/CD pipeline moves that detection to the point where it’s cheapest to fix: before the code ships. Every pull request that touches agent logic runs your configured attack scenarios, and a failing eval blocks the deploy just like a failing unit test.

​
Install tridentctl

Install the Trident CLI in your CI environment before running evaluations:
If you prefer not to install globally, you can invoke it directly with npx:

​
Set your project credentials

Trident authenticates eval runs using your project API keys. Add the following as secrets or environment variables in your CI system: Never hard-code these values in your repository. Use your CI system’s secrets manager to inject them at runtime.

​
Run evaluations

1

Create an eval configuration file

Write a YAML file (for example, trident-evals.yaml) using the promptfoo configuration format. Define the prompts, attack scenarios, and assertions you want to run. Start with high-priority attacks like prompt injection and jailbreak attempts.
2

Run the eval command

Execute the following command in your CI pipeline:
Trident runs each scenario in the eval corpus against your agent and evaluates the responses using the assertions defined in your config file.
3

Interpret the exit code

The tridentctl evals command exits with a non-zero status code if any test assertion fails. Most CI systems treat a non-zero exit as a pipeline failure, which blocks the pull request or deployment automatically — no additional configuration required.
4

Use --watch for local development (optional)

During local development, pass the --watch flag to have tridentctl automatically re-run evaluations whenever your config file or prompt files change:

​
GitHub Actions example

Add the following workflow file to your repository to run Trident evals on every pull request:
.github/workflows/trident.yml
The workflow runs on every pull request. If any eval fails, the check fails and the PR cannot be merged (assuming branch protection is enabled on your default branch).

​
Trigger evaluations via API

If you want to trigger a full cloud-run campaign rather than running evals locally in the CI runner, you can call the Trident REST API instead: Start a campaign:
Poll for completion:
Poll until the response status field is completed or failed, then use the passed boolean to determine whether to proceed with the deploy.
Start your CI gate with a small, focused set of high-priority attacks — prompt injection, jailbreak, and any domain-specific risks your agent faces. Once those pass consistently, expand the corpus incrementally. A gate that runs in under two minutes is far more likely to stay in place than one that takes ten.