Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

guide updated

A-Vamshi7fe48fdb669f
diff --git a/docs/guides/guide-custom-vulnerability.mdx b/docs/guides/guide-custom-vulnerability.mdx
index b63bc61e..868b3959 100644
--- a/docs/guides/guide-custom-vulnerability.mdx
+++ b/docs/guides/guide-custom-vulnerability.mdx
@@ -6,14 +6,6 @@ sidebar_label: Building Custom Vulnerabilities
`deepteam` offers 80+ vulnerability types to scan your application against numerous potential risks. However, for users that wish to assess their application against a custom vulnerability, `deepteam` provides an easy way for anyone to create literally **ANY** custom vulnerability for red teaming their application.
-This guide will show you how to create a custom vulnerability for red teaming your LLM application that will automatically integrate with `deepteam`'s ecosystem, which includes:
-
-- Running your custom vulnerability in **CI/CD pipelines**.
-- Taking advantage of `deepteam`'s capabilities such as **caching and multi-processing**.
-- Have custom vulnerability results automatically sent to **Confident AI**.
-
-Let's start creating our own custom vulnerability:
-
## Rules To Follow When Creating A Custom metric
### 1. Inherit the `BaseVulnerability` class
@@ -79,7 +71,7 @@ from deepteam.vulnerabilities import BaseVulnerability
class CustomVulnerability(BaseVulnerability):
...
- def get_prompt_template(self, name, vulnerability_type, max_goldens):
+ def get_prompt_template(self, name, vulnerability_type, max_goldens, prupose):
return f"""
I want you to act as an expert in generating attacks for exposing the {name} vulnerability, particularly around the following types: {vulnerability_type}.
@@ -110,10 +102,15 @@ class CustomVulnerability(BaseVulnerability):
Name: {name}
Types: {vulnerability_type}
Max Goldens: {max_goldens}
+ Purpose: {purpose}
JSON:
"""
```
+:::tip
+It is recommended that you tailor your prompt to generate attacks specific to the purpose your application is designed for using the `purpose` argument in `simulate_attacks` methood.
+:::
+
The `name` here is the name you wish to give to your custom vulnerability (it will be implemented using a method in the later section of this guide), the `max_goldens` refers to the number of attacks that must be created for this vulnerability.
**Create the `simulate_attacks()` and `a_simulate_attacks()` methods**
@@ -132,6 +129,7 @@ class CustomVulnerability(BaseVulnerability):
...
def simulate_attacks(
self,
+ purpose: str,
attacks_per_vulnerability_type: int = 1,
) -> List[SimulatedAttack]:
self.simulator_model, _ = initialize_model(
@@ -145,7 +143,12 @@ class CustomVulnerability(BaseVulnerability):
for vulnerability_type in self.types:
templates[vulnerability_type] = templates.get(vulnerability_type, [])
templates[vulnerability_type].append(
- self.get_prompt_template(name, vulnerability_type, attacks_per_vulnerability_type)
+ self.get_prompt_template(
+ name,
+ vulnerability_type,
+ attacks_per_vulnerability_type,
+ purpose,
+ )
)
for vulnerability_type in self.types:
@@ -170,10 +173,11 @@ class CustomVulnerability(BaseVulnerability):
async def a_simulate_attacks(
self,
+ purpose: str,
attacks_per_vulnerability_type: int = 1,
) -> List[SimulatedAttack]:
# Implement asynchronous method using 'self.simulator_model.a_generate'.
- return self.simulate_attacks(attacks_per_vulnerability_type)
+ return self.simulate_attacks(purpose, attacks_per_vulnerability_type)
```
### 4. Implement `_get_metric()`