DeepTeam and promptmap2: Next-Gen LLM Red Teaming Frameworks#

Core Capabilities#

DeepTeam provides a structured approach to LLM security testing across three pillars:

1. Vulnerability Detection

With 50+ state-of-the-art, ready-to-use vulnerabilities, DeepTeam covers a broad spectrum of LLM security risks:

2. Attack Simulation

DeepTeam implements multiple attack strategies including:

• Gray-box attacks: Testing with partial knowledge of the system
• Prompt injection: Direct and indirect injection techniques
• Jailbreak attempts: Multi-turn jailbreak strategies
• Social engineering: Manipulative prompts designed to bypass safety filters

3. Guardrail Implementation

After identifying vulnerabilities, DeepTeam helps add guardrails to secure your LLM systems:

1
from deepteam import red_team
2
from deepteam.vulnerabilities import PromptInjection, PIILeakage
3
from deepteam.attacks import GrayBoxAttack
4

5
# Define your LLM system
6
def my_llm_app(input_text):
7
    # Your LLM application logic
8
    return response
9

10
# Run red team assessment
11
results = red_team(
12
    target=my_llm_app,
13
    vulnerabilities=[PromptInjection(), PIILeakage()],
14
    attacks=[GrayBoxAttack()]
15
)

Framework Integration#

DeepTeam supports structured testing aligned with major security frameworks:

• OWASP Top 10 for LLMs: Maps findings to the OWASP LLM risk categories
• NIST AI RMF: Aligns with the NIST risk management framework
• Custom frameworks: Define your own testing methodology

How DeepTeam Differs from garak and promptfoo#

What promptmap2 Does#

promptmap2 takes a systematic approach to prompt injection testing:

1. System prompt analysis: Examines your LLM’s system prompt for potential weaknesses
2. Automated attack generation: Creates targeted injection payloads based on the system prompt
3. Multi-vector testing: Tests injection through multiple input channels
4. Impact assessment: Evaluates the severity of successful injections

Key Improvements Over promptmap v1#

• Broader attack coverage: More injection techniques and mutation strategies
• Better false positive handling: Improved detection of actual vs. apparent injection success
• Custom application support: Designed specifically for custom LLM applications, not just raw models
• Automated reporting: Structured output for integration with security workflows

Use Cases#

promptmap2 is particularly valuable for:

• Custom LLM applications: Test your specific application logic, not just the base model
• System prompt hardening: Identify and fix weaknesses in your system prompts
• Pre-deployment testing: Run prompt injection tests before releasing LLM applications
• Continuous security: Integrate into CI/CD for ongoing prompt injection regression testing

Comparison with Other Injection Scanners#

Recommended Testing Pipeline#

1
Phase 1: Broad Scanning
2
├── garak — Comprehensive security probing
3
├── DeepTeam — Safety + security + guardrails
4
└── promptfoo — Prompt regression testing
5

6
Phase 2: Targeted Testing
7
├── promptmap2 — System prompt injection testing
8
├── whistleblower — Prompt leakage assessment
9
└── FuzzyAI — Jailbreak fuzzing
10

11
Phase 3: Custom Attacks
12
├── PyRIT — Microsoft's risk identification
13
├── DeepTeam custom vulnerabilities
14
└── Manual red team exercises

CI/CD Integration#

Both tools can be integrated into development workflows:

1
# GitHub Actions example
2
- name: LLM Red Team
3
  run: |
4
    pip install deepteam promptmap2
5
    deepteam scan --target my_llm_app --output results.json
6
    promptmap2 --endpoint http://localhost:8000 --system-prompt ./system.md