evaluation_id. Use vijil evaluate to start the evaluation, then use vijil evaluations redteam-* to inspect it.
Start a Red Team Evaluation
Run a Red Team evaluation and wait for a terminal status:succeeded, failed, cancelled, and degraded. A degraded evaluation finished with some attacker failures; inspect its status details before using the report.
Add --no-wait to return immediately after creation. Save the evaluation ID from the output for later commands:
Configure the Evaluation
Pass a JSON object through--redteam to override the default wave settings:
Red Team runs directly against the registered Agent. It does not accept
--baseline, --bespoke, --sample-size, --local, or --harness-name.
Inspect Progress and Evidence
Each command accepts
--team <team-id>. If you omit it, the CLI uses the default team from your configuration.
Status
Show the evaluation’s status, current phase, wave progress, attacker counts, errors, and degradation details:Seeds
Seeds describe the risks and instructions that attackers explore. List every seed or filter by wave:Attacks
Attack results include the attacker type, execution status, transcript, and error details. List results by evaluation or wave:--wave and --seed:
Judgments
Judgments record whether an attack succeeded, whether it could cause harm, and whether the response leaked artifacts:--wave and --seed:
Reflections
After each completed wave, the evaluation summarizes successful seeds, saturated outcomes, and new strategy patterns. List all reflections or select one wave:Retrieve the Report
The final report summarizes vulnerabilities, policy violations, leaked artifacts, and successful strategies. JSON is the default format:--format. Use --output to write the report to a file; PDF output requires it:
--force-regenerate to rebuild a cached PDF:
For the shared behavioral and Red Team startup options, see the CLI evaluation guide.
Red Team Engagements
A Red Team Engagement runs adversarial attacks against your Agent in waves, learning from each wave to plan the next. See Engagements for how the loop works.vijil redteam create
Create and start an Engagement. Only --agent-id is required, because the service fetches the target endpoint and Agent card from the Agent registry.
Save the Engagement ID from the output:
--redteam controls the shape of the run, such as min_waves, max_waves, max_seeds_per_wave, and max_parallel_attackers. --advanced controls how hard each attacker works before giving up. Both map onto the Console sliders described in Run a Red Team Campaign. The exact JSON keys come from the vijil-redteam schemas package, so check vijil redteam create --help for the set your deployment accepts.vijil redteam list
List Engagements for the active team, showing engagement_id, status, phase, and created_at.
vijil redteam status
Get one Engagement’s live state, including phase, current wave, attackers completed, and attackers errored.
vijil redteam results
Get the aggregated results for an Engagement.
vijil redteam cancel
Cancel a running Engagement.
vijil redteam delete
Delete an Engagement.
vijil redteam seeds
List the attack seeds generated for an Engagement, showing wave_id, seed_id, instruction, and risk_type.
vijil redteam attacks
List attacker runs, showing wave_id, seed_id, attacker_type, and status.
vijil redteam attack
Get one attacker run, including its strategy and transcript.
vijil redteam judgments
List judgments, showing wave_id, seed_id, is_success, and potential_harm.