Skip to main content
Run a Red Teaming test, fix what breaks, and run it again, so you end up with a passing report to share with your security or compliance reviewer.

​
Before you start

  • A voice, chat, or SMS agent in Bluejay. Red Teaming does not support email agents. See Connect your agent.
  • For Custom intensity, the Growth, Scale, or Enterprise plan.

​
Steps

1

Launch a red team

Open Red Teaming and click Launch red team. Choose the target agent, then an Intensity.The Launch red team dialog with a target agent selected and the Quick, Standard, Thorough, and Custom intensity cardsThe Launch red team dialog with a target agent selected and the Quick, Standard, Thorough, and Custom intensity cards
  • Use Quick for a smoke test while you iterate, and Thorough for the run you share, because it escalates deepest.
  • Custom sets the number of attempts for each attack family, and a family set to 0 is skipped. Launch stays disabled until at least one family is above 0.
2

Give the attacker partial account details

To test your verification steps, open the Attacker knowledge tab and add what a real attacker might already know about the customer, for example a date of birth or the last four digits of a card. Only the social engineering callers use it.The Attacker knowledge tab with one trait, Date of Birth, filled inThe Attacker knowledge tab with one trait, Date of Birth, filled inIf the attackers dial into an inbound agent that needs an extension, enter it on the Connection tab.
3

Start the run and wait for it to finish

Click Launch red team. The callers first learn what data and tools your agent has, then push harder with the tactics that work. The run ends when it reaches its call limit, or when you stop it.The Swarm tab showing the target agent and the attackers, with breached and held attacks and one still evaluatingThe Swarm tab showing the target agent and the attackers, with breached and held attacks and one still evaluating
4

Read the verdict for each category

On the Report tab, start with the verdict and the categories listed under Breached. Then check Conceded ground, where the agent resisted but gave up some ground. Categories under Not assessed did not run, so a higher intensity is the way to cover them. See Read the results for what each verdict means.A Red Teaming report with the Breached verdict, attack and breach counts, and the OWASP categories that were breached and not assessedA Red Teaming report with the Breached verdict, attack and breach counts, and the OWASP categories that were breached and not assessedContent-harm categories are in the same report. A content-harm failure alone fails the run. See OWASP coverage and content harm.
5

Open each breach

On the Swarm tab, click an attacker marked Breached, then open Transcript to read the conversation that breached the agent. Read the words that worked and find the gate that was missing or bypassed.The Transcript tab of a breached attack, with the recording player and the conversation between the agent and the attackerThe Transcript tab of a breached attack, with the recording player and the conversation between the agent and the attacker
6

Fix each breach

Pick the fix that matches the gap:
  • The agent said something it should not have: tighten its prompt.
  • The agent acted without checking who was calling: add or tighten the verification step.
  • The agent took an action the caller should not reach: remove tool access the agent does not need.
7

Run the same intensity again

Launch again with the same intensity and attacker knowledge, so the two reports are comparable. Changing either one makes a better score hard to trust.Two completed Red Teaming runs for the same agent, a later run that passed and an earlier run that failedTwo completed Red Teaming runs for the same agent, a later run that passed and an earlier run that failed

​
Check the result

The fix worked when the categories that showed FAIL now show PASS. Held under pressure means the agent resisted but conceded ground on some attacks, so read those attacks before you call the agent ready. If a category is still breached, go back to the conversation, because the fix may have closed one path and left another. When the report is clean, click Download PDF to save a copy for your reviewer. Download CSV gives the same results as a spreadsheet.
Last modified on October 22, 2026