Use case: Red Teaming

Break prompts on purpose. Fix them with evidence.

You're crafting injection attacks, testing guardrail bypasses, and documenting defense iterations. You need version-controlled attack libraries, multi-model execution, and a safety scanner you can run on both sides. Not another chat window.

How Promptmark fits

Four-layer safety scanning

Run the same scanner your targets use. PII detection, injection pattern matching, secrets scanning, and AI-powered moderation — test your attacks against all four layers. Validate that your defenses catch what they should and document what slips through.

Template variables for attack variants

Parameterize your adversarial prompts: {{injection_type:select:direct,indirect,recursive}}, {{payload:text}}, {{target_behavior:text}}. One attack template generates a taxonomy of variants. Schema validation keeps your test matrix consistent across runs.

Multi-model testing

The same attack hits different models differently. Run adversarial prompts against 300+ models and compare which ones resist, which ones comply, and which ones hallucinate around the guardrails. Document model-specific vulnerabilities with real data.

Playbooks for automated test sequences

Chain multi-step attack sequences into playbooks. Escalate from reconnaissance to injection to exfiltration across steps. Branch on model responses. Run the full sequence against new models or prompt versions with one trigger URL.

Version-controlled attack iterations

Every edit to an attack prompt is saved automatically. Diff two versions to see exactly how an injection evolved.

Collections for attack taxonomies

Organize attack prompts by OWASP category, model target, injection type, or severity. Tag across collections to find every prompt-leaking variant.

Share findings with your security team

Publish sanitized attack collections to your profile or share via direct link. Team members see the exact prompt versions and test results.

Example workflow

1

Build the attack library

Create prompt templates for each attack category: direct injection, indirect injection, jailbreaks, prompt leaking, role manipulation. Add template variables for payload variants. Organize by attack taxonomy using collections and tags.

2

Test across models

Run each attack template against target models with controlled inputs. Compare which models resist, which comply, and how responses differ. Save results alongside the exact prompt version and parameters used.

3

Validate defenses

Write defensive system prompts and run the same attacks against them. Use the safety scanner to check both your attacks and your defenses. Version control tracks every iteration of the attack-defense cycle.

4

Automate with playbooks

Build playbooks that chain attack sequences: initial probe, escalation, payload delivery, exfiltration attempt. Run the full chain against new model versions or updated defenses. Deliver results to your security team via webhook.

5

Document and share findings

Organize attack collections by vulnerability class. Version history shows how attacks evolved and which defense iterations closed each gap. Publish sanitized findings to your profile or share collections with your security team.

Build a prompt security lab, not a folder of text files

Version-controlled attack libraries, parameterized variants, multi-model execution, and built-in safety scanning. Red team with real tools.

Set up your testing environment — free