Run tester attacks

Testers pick a Task and Objective in the Annotation Queue, attack the model in Chat Playground, and submit the results.

1. Select a Task and Objective

①Annotation Queue

Testers pick a Task to join from Annotation Queue in the left menu. Testers see only Active Tasks.

Progress status	Meaning
`Waiting`	A Task with no submissions yet
`In Progress`	A Task with at least one Submission
`Complete`	A Task where every Objective has at least one submission

A Submission is one attack attempt by a tester.
When you open a Task, the top shows the target model, your submission count, your success count, and the number of Objectives you succeeded on.
The Objective list below shows, for each Objective, other testers' attack success count, the Judge Type, the Judge Model, and whether you've submitted.

1-1Pick a Task
1. Click Annotation Queue in the left menu.
2. Check the Task's progress status.
3. Click the Task you want to join.
Closed Tasks don't appear in the list.
1-2Pick an Objective
1. Check the goals and descriptions in the Objective list.
2. Check other testers' attack success counts and whether you've submitted.
3. Click the Objective to attack to open Chat Playground.

2. Attack the model

①Objective Info

Check the attack conditions in Objective Info on the left of Chat Playground.

Field	Meaning
`Note To Tester`	Useful info for testers (e.g., good seeds to start with, examples of successful attacks)
`Target Model`	The model under attack
`Judge Type` · `Judge Model`	The scoring method and scoring model
Rubric · `Attack Success Threshold`	The scoring criteria and success condition. Visible only when the admin turns on `Show Rubric to Testers`.

②Multi-turn attacks

You attack through a multi-turn conversation with the model in Chat in the center. Even if you're refused once, you can change the context and keep trying. When you submit, the whole conversation is recorded.

2-1Check the attack conditions
1. Check the goal in Note To Tester.
2. Check the Target Model and Judge settings.
3. If the Rubric is visible, check the success condition.
If you don't see a Rubric, the admin hasn't made it visible for this Objective.
2-2Attack through conversation
1. Send a message in the input box to start the attack.
2. Keep asking based on the responses.
See Multi-turn attacks.

3. Submit results and check the Verdict

①Submit

When you finish the attack, submit it in the Submit panel on the right.

If the Judge Type is Self Evaluation, answer Did your attack succeed? with Succeeded or Failed.
Comment is optional. Note your attack method or anything worth knowing.
If there's no conversation, Submit is disabled.
There's no limit on the number of submissions. You can try again until the Task is closed.

②Verdict

A Verdict is the final judgment of a submission. When you submit, you move to the Submissions tab and see the scoring result.

Item	Display
Final judgment	`Attack Succeeded` (red) / `Attack Blocked` (green)
Result per Rubric item	`Criteria Met` (red ✓) / `Criteria Not Met` (gray ✕)
`Judge Reason`	The Judge's reasoning for its judgment
`View Transcript`	The full submitted conversation

3-1Submit results
1. For Self Evaluation, select Succeeded or Failed.
2. Enter a Comment if needed.
3. Click Submit.
There's no limit on the number of submissions.
3-2Check the Verdict
1. See your submissions in the Submissions tab.
2. Check the final judgment.
3. Check the result per Rubric item and the Judge Reason.
4. See the full conversation with View Transcript.
See Verdict for what each display means.

Frequently asked questions

Why can't I click Submit?Because there's no conversation yet.

Cause: A submission saves the conversation history, so there must be at least one message.
Solution: Talk to the model in Chat first, then submit.

Why can't I see the Rubric and Attack Success Threshold?Because the admin hasn't made them visible.

Cause: For Objectives with Show Rubric to Testers turned off, the scoring criteria aren't shown to testers.
Solution: Attack using the Note To Tester without the criteria. If needed, ask the admin to make them visible.

Why did the Task I was working on disappear from the list?Because the admin closed the Task.

Cause: Closed Tasks don't appear in testers' Annotation Queue and don't accept new submissions.
Solution: Check the Task status with your admin.

Can I submit more than once for the same Objective?Yes, there's no limit.

Cause: Each submission is recorded as a separate Submission and scored on its own.
Solution: Attack again with a different approach and submit. Results accumulate in the Submissions tab.