Skip to content

Knox RT Red Teaming Policies (Beta)¶

Beta

Red Teaming Policies is in beta. The screens on this page come from the design preview and can change before release.

Knox RT (AccuKnox Red Teaming) tests your AI models and agents with a red teaming policy. A policy holds the template, the prompts, the attack strategies, the targets and the schedule. You build it once, then run it on demand or on a cron.

Every Policy Sits in One List¶

Open Red Teaming Policies. Each row shows the models and agents it targets, its category and plugin coverage, and its findings by severity.

Red Teaming Policies list with the tabs, Create Policy button, coverage and findings columns

  • Tabs split the list into All, Models and Agents.
  • Filter narrows the list by model, agent or category. Saved Filter brings back a view in one click.
  • System policies ship with AccuKnox. User policies are the ones your team builds.

Build a Policy in Four Steps¶

Click Create Policy. The drawer walks you through four steps.

1. Policy Configuration¶

Name the policy and pick a starting point.

  • Predefined Template starts from an AccuKnox template.
  • Create New Template starts from the categories you pick. See Build your own template.

2. Modules and Plugins¶

Pick a template. Each card shows its module and plugin count.

Template cards and the three prompt sources in step 2

Template What it covers
OWASP LLM Top 10 The OWASP Top 10 risks for LLM applications
EU AI Act Readiness Risks relevant to EU AI Act conformity assessments
Harmful Content & Safety Harmful, criminal and biased output across personas
Deep Adversarial Multi-turn, iterative attacks on guardrails and agents
Privacy & Access Personal data disclosure and authorization bypass
Baseline Coverage Every category at low depth

Then pick a prompt source.

  • Default Prompts uses the prompt library for each module. You change only severity and which plugins run.
  • Customize Default Prompts lets you set test counts, language and per-module overrides.
  • Upload your custom prompts takes your own prompt set as CSV, JSON or YAML.

Turn modules on or off, then pick the Attack Strategies.

Attack strategies with base64, jailbreak, fiction and multi-turn crescendo

Crescendo is a multi-turn strategy. It escalates over several turns, the way a patient attacker does. Click the gear icon to set its turn limits.

3. Review Configuration¶

Check the policy summary, the generated prompts and the policy YAML before you save.

4. Targets and Schedule¶

Targets and Schedule step with the model and agent pickers and the cron fields

  1. Pick Target Models and Target Agents. Targets are optional. You can save now and attach them later from the policy details.
  2. Pick On-demand, or Scheduled and fill in the cron fields. The drawer shows the next run in server time (UTC) and in your time zone.
  3. Click Save Policy.

Build Your Own Template¶

Create New Template builds a template from the categories you pick, with prompts generated for each plugin.

  • Plugin override. Give any plugin its own instructions, attack strategies and multi-turn limits.
  • Custom plugin. Add a plugin with its own objective that fits your app, such as "bypass the refund limit".

Edit, Clone and Delete Policies¶

Open the row menu on any policy.

Action What it does
Edit Changes the policy. The name and method stay locked, because run history and findings link to them.
Clone Copies the policy under a new name.
Delete Removes the policy and its schedule. Findings from earlier runs stay in Findings.

Select several rows to clone or delete them together.

One Drawer Holds Every Result¶

Click a policy name to open its detail drawer.

Policy detail drawer with tabs, run results and the run, edit, clone and delete icons

Tab What you see
Overview Attack success rate, pass rate, test cases, average duration, schedule, targets, strategies and coverage by category
Findings Each failed attack, with its category, severity, model and time
Runs The findings trend across runs, and each run's status
Audit Trail Every change to the policy and who made it
YAML The policy as YAML, to copy or download

Results Land in AccuKnox Findings¶

Go to Issues > Findings and pick AI Red Teaming Findings. Each row holds the attack prompt, the model output, the strategy and the reason it failed.

AI Red Teaming Findings list with the attack prompt, output, strategy and reason columns

Findings map to the OWASP Top 10 for LLM and MITRE ATLAS. Rules, tickets and exports work the same as for any other finding.

Start with a template, then narrow it

Run Baseline Coverage first to see where a model fails. Then build a focused policy, such as Deep Adversarial, for the categories that failed.

Related pages