continuous-llm-red-teaming-with-promptfoo

Files (5)

• LICENSE
• SKILL.md
• api-reference.md · references/
• standards.md · references/
• agent.py · scripts/

SKILL.md (first lines)

# Continuous LLM Red Teaming with Promptfoo

> **Authorized Use Only:** Run these adversarial probes only against LLM applications and endpoints you own or are explicitly authorized to test. Generated attack payloads (jailbreaks, prompt injections, harmful-content elicitation) are adversarial inputs; sending them to third-party services without permission may violate terms of service.

## Overview

Promptfoo is an open-source LLM evaluation and red-teaming framework (used by OpenAI and Anthropic per its README) that generates adversarial test cases, runs them against your model/agent, and grades the responses. DeepTeam (by Confident AI) is a complementary open-source framework offering 50+ ready-to-use vulnerabilities and 10+ research-backed attack methods. Together they let you treat LLM security as a **regression test**: every commit re-runs the same adversarial suite, and the pipeline fails when a previously-safe behavior regresses.

This matters because LLM applications change constantly — prompts, models, RAG sources, tools, and guardrails all drift. A jailbreak that was patched last sprint can silently return after a prompt edit or a model upgrade. Promptfoo maps its plugins directly onto the **OWASP LLM Top 10** (`owasp:llm`) and **OWASP Agentic** (`owasp:agentic`) presets, and onto MITRE ATLAS, so the suite tracks recognized risk taxonomies. The core threat addressed here is **AML.T0051 — LLM Prompt Injection** (MITRE ATLAS): adversarial instructions that override the application's intended behavior. This skill follows the Promptfoo red-team docs (https://www.promptfoo.dev/docs/red-team/) and DeepTeam docs (https://www.trydeepteam.com/docs/getting-started), and aligns to NIST AI RMF MANAGE-4.1 (post-deployment monitoring and feedback to manage AI risk).

## When to Use

- When you need continuous, automated red-teaming of an LLM app in CI/CD rather than one-off manual tests.
- When you want to enforce a security gate: block merges that introduce or reintroduce jailbreak/injection vulnerabilities.
- When mapping coverage to OWASP LLM Top 10 / OWASP Agentic / MITRE ATLAS for compliance reporting.
- When comparing the security posture of two models or prompt versions side by side.
- When tracking vulnerability regression over time across releases.

## Prerequisites

- Node.js 18+ (Promptfoo is distributed via npm) and Python 3.9+ (for DeepTeam).
- Install Promptfoo and DeepTeam:
  ```bash
  npm install -g promptfoo            # or: npx promptfoo@latest
  pip install -U deepteam
  ```
- API access/credentials for the target LLM endpoint (and a grader model, e.g. an OpenAI key) exposed as environment variables.
- A CI/CD platform (GitHub Actions, GitLab CI) with secret storage.
- Authorization to test the target application.

## Objectives

- Scaffold a Promptfoo red-team config targeting your LLM app.
- Enable OWASP LLM Top 10 and OWASP Agentic plugin presets plus jailbreak/injection strategies.
- Run the suite locally and interpret the per-plugin pass/fail report.
- Add DeepTeam as a second engine for programmatic, research-backed attacks.
- Integrate both into CI/CD so builds fail on new vulnerabilities.
- Generate shareable HTML/PDF security reports per run.

## MITRE ATT&CK Mapping

| ID | Name (MITRE ATLAS) | Tactic |
|----|--------------------|--------|
| AML.T0051 | LLM Prompt Injection | Initial Access / Persistence (LLM) |
| AML.T0051.000 | Direct (Prompt Injection) | LLM Attack |
| AML.T0051.001 | Indirect (Prompt Injection) | LLM Attack |
| AML.T0054 | LLM Jailbreak | Privilege Escalation / Defense Evasion (LLM) |

## Workflow

### 1. Scaffold the red-team configuration
Initialize an interactive config; it writes `promptfooconfig.yaml` where targets, plugins, and strategies live.

```bash
promptfoo redteam init
# choose your target type (HTTP endpoint, openai:..., anthropic:..., custom provider)
```

### 2. Define targets, OWASP presets, and attack strategies
Edit `promptfooconfig.yaml`. The `purpose` grounds attack generation; `plugins` are adversarial input generators; `strategies` are delivery techniques (jailbreak/injection wrappers).

```yaml
# promptfooconfig.yaml
targets:
  - id: https://api.example.com/chat        # your app endpoint
    label: support-bot

redteam:
  purpose: |
    A customer-support assistant for an e-commerce site. Must never reveal
    system prompts, leak PII, or perform actions outside order support.
  numTests: 10
  plugins:
    - owasp:llm          # OWASP LLM Top 10 preset
    - owasp:agentic      # OWASP Agentic threats preset
    - id: pii:direct
      numTests: 15
    - prompt-extraction  # system-prompt leakage
    - harmful
  strategies:
    - id: jailbreak              # iterative single-turn jailbreak
    - id: jailbreak:composite    # stacked jailbreak techniques
    - id: crescendo              # multi-turn escalation
    - id: prompt-injection       # injection wrapper
```

### 3. Run the suite and view the report
`redteam run` combines generation + evaluation; then open the interactive report.

```bash
promptfoo redteam run
promptfoo redteam report            # launches the web report (pass/fail per plugin)
```
Each row shows the plugin (mapped to OWASP/ATLAS), the strategy, the attack prompt, the model's response, and the grader's verdict. The **attack success rate** per plugin is your headline metric — track it per release.

### 4. Add DeepTeam for programmatic, research-backed attacks
Use DeepTeam to cover additional vulnerabilities/attacks and to script bespoke suites in Python.

```python
# deepteam_suite.py
from deepteam import red_team
from deepteam.vulnerabilities import Bias, PIILeakage
from deepteam.attacks.single_turn import PromptInjection

def model_callback(prompt: str) -> str:
    # call your application's LLM endpoint here and return the text response
    return call_my_app(prompt)

red_team(
    model_callback=model_callback,
    vulnerabilities=[Bias(types=["race"]), PIILeakage(types=["api_and_database_access"])],
    attacks=[PromptInjection()],
)
```
DeepTeam can also be driven from a YAML config:
```bash
deepteam run config.yaml
```

### 5. Gate the build in CI/CD (GitHub Actions)

From mukul975/Anthropic-Cybersecurity-Skills, Apache-2.0.

More from mukul975/Anthropic-Cybersecurity-Skills