promptfoo-redteam-run

Promptfoo Redteam Run

Execute redteam probes reproducibly, inspect the output artifact, and rerun only the slice that needs attention. Prefer evaluating existing generated tests with redteam eval; regenerate with redteam run only when config/test generation must change.

Read references/redteam-run-patterns.md when you need command recipes, result inspection snippets, or CI examples.

Inputs

Infer these from the repo or user prompt:

• Generated scan file, usually redteam.yaml, or setup config if regeneration is requested.
• Target environment, secrets/env file, concurrency/rate limits, and whether cloud sharing is allowed.
• Grading mode: default remote grading, redteam.provider, --grader, or local deterministic QA provider.
• Desired output: JSON/YAML/HTML export, report, CI gate, failure triage, or rerun of a previous result.

If a target or generated tests are missing, use promptfoo-provider-setup or promptfoo-redteam-setup first.

Workflow

1. Choose run mode

• Use redteam eval when redteam.yaml already exists and you want stable apples-to-apples runs.
• Use redteam run --force only when the setup changed or the user wants fresh generated probes.
• Use a configured redteam.provider or --grader only when the scan needs a specific generator/grader or deterministic QA behavior.
• Disable cloud sharing by default for internal targets. redteam eval and retry accept --no-share; redteam run does not currently expose that flag, so export PROMPTFOO_DISABLE_SHARING=true for the whole invocation or split into redteam generate + redteam eval --no-share. Only re-enable sharing when the user explicitly asks for a cloud URL.

2. Preflight

Use the CLI form that matches your environment for every command below: from the promptfoo repo, npm run local -- redteam … (align Node first with source ~/.nvm/nvm.sh && nvm use); outside the repo (an installed plugin or your own app project), npx promptfoo@latest redteam …, or a globally installed promptfoo redteam ….

Validate first:

npm run local -- validate config -c path/to/redteam.yaml
npm run local -- validate target -c path/to/redteam.yaml
# Outside the repo:
npx promptfoo@latest validate config -c path/to/redteam.yaml

Check that generated tests include assert, metadata.pluginId, metadata.purpose or defaultTest.metadata.purpose, and the real input vars. For file:// providers in a generated eval file, target providers resolve like normal config file providers, so file://./target.mjs is relative to that config file. redteam.provider is loaded during grading from the command working directory, so use an absolute path or a repo-root-relative path when running from the repo root.

If validation fails with ENOENT for file://./target.mjs, the generated YAML was probably written to a different directory than the target. Regenerate beside the source config, move the generated file next to the target, or change the target id to an absolute/repo-root-relative file:// path before rerunning.

3. Run and export

Evaluate generated tests:

npm run local -- redteam eval -c path/to/redteam.yaml -o /tmp/redteam-results.json --no-cache --no-share --no-progress-bar
# Outside the repo:
npx promptfoo@latest redteam eval -c path/to/redteam.yaml -o /tmp/redteam-results.json --no-cache --no-share --no-progress-bar

Generate and evaluate in one command only when needed. redteam run has no --no-share flag, so disable sharing via the environment variable:

PROMPTFOO_DISABLE_SHARING=true npm run local -- redteam run -c path/to/promptfooconfig.yaml --force --no-cache --no-progress-bar

For fragile targets, set -j 1 and add --delay rather than allowing broad concurrency.

4. Inspect results

Always inspect the exported artifact. Do not rely only on the exit code because redteam failures may intentionally return a failing exit status.

Look for:

• results.stats.successes, failures, errors, and tokenUsage
• Failed or errored rows, including response.output, error, gradingResult, metadata.pluginId, metadata.strategyId, and target label
• shareableUrl; it should be null when --no-share is used
• Attack success rate: failures / (successes + failures)

Treat grader transport/parse failures separately from real target failures. If --filter-errors-only returns zero rows, the source result likely had no ERROR rows or the generated test indices changed since the source run.

5. Rerun narrowly

Use filters before rerunning expensive scans:

npm run local -- redteam eval -c path/to/redteam.yaml --filter-failing /tmp/redteam-results.json -o /tmp/redteam-failing-rerun.json --no-cache --no-share --no-progress-bar
npm run local -- redteam eval -c path/to/redteam.yaml --filter-errors-only /tmp/redteam-results.json -o /tmp/redteam-errors-rerun.json --no-cache --no-share --no-progress-bar
npm run local -- redteam eval -c path/to/redteam.yaml --filter-metadata pluginId=policy -o /tmp/redteam-policy.json --no-cache --no-share --no-progress-bar

For error-only reruns that should update the original evaluation in place, use promptfoo retry <evalId> instead of creating another eval.

6. Report or gate

Use redteam report for interactive triage after results are written. It starts or reuses the local Promptfoo UI, so ask before running it unless the user explicitly requested the report UI:

npm run local -- redteam report

For CI, gate on explicit metrics from the JSON export. Keep thresholds tied to the app's risk tolerance and track category-level changes with metadata.pluginId.

Common Mistakes

# WRONG: regenerates probes when you only wanted a comparable rerun
promptfoo redteam run

# BETTER: reuse generated tests
promptfoo redteam eval -c redteam.yaml -o results.json --no-cache --no-share

# WRONG: broad rerun after a flaky target error
promptfoo redteam eval -c redteam.yaml

# BETTER: rerun only target/grader errors from the prior result
promptfoo redteam eval -c redteam.yaml --filter-errors-only results.json -o errors-rerun.json --no-cache --no-share

Output Contract

When done, state:

• Run mode used: existing generated tests or regenerate-and-run
• Config/result/report paths
• Data-sharing mode and grader/provider used
• Commands run and whether any returned nonzero due to redteam failures
• Success, failure, error counts and attack success rate
• Highest-priority failing plugins/strategies and recommended next rerun or fix