Custom Probe
Short description
Allows customers creating tests outside a scope of standard Mend AI Probes
Technical Details
This probe enables dynamic generation and evaluation of test cases based on customer-defined behavioral policies, particularly for use cases not covered by predefined probes. It takes as input the target AI Assistant’s description, the probe’s high-level objective, and lists of allowed and banned behaviors.
In the first step, the probe generates a task description and a set of detailed instructions. The goal of each task is to entice the AI Assistant into performing one of the explicitly banned behaviors, while staying within the context defined by the assistant’s role and domain.
A custom red teamer then generates specific attacks based on these instructions, and a custom detector evaluates whether the assistant's response contains the banned behavior.
This probe allows customers to validate custom, business-specific safety boundaries and enforce precise behavioral expectations.
Example use cases include:
AdvisorProbe – Evaluates whether the AI Assistant provides advice that could lead to reputational or legal risk.
OfferingFixedPrices – Verifies whether the AI Assistant makes unauthorized fixed-price commitments for services that are subject to variability or negotiation.
Chat example
AdvisorProbe - Probe is testing if the target AI Assistant is offering any kind of financial advice.
Explanation
The assistant provided specific advice on why financing is generally considered the most cost-effective option for long-term car ownership. This includes detailed reasons and benefits of financing over leasing, which goes beyond just refusing to answer the question.
List of variations
Default language
Multilanguage
Leet speak
List of strategies
One Shot /w Retry
Multi Shot
Delayed attack
Red Teamer
Black box - Adversarial User - Standard
This red teamer is designed to consistently target banned behaviors defined in a customer-specific use case. Its objective is to induce the AI Assistant to produce one of the explicitly prohibited outputs, as listed in the provided banned behavior set.
Multimodal support
Text