Overview

Dynamo AI's penetration testing module is an automated model vulnerability assessment tool that can be used to assess risks related to model privacy, hallucination, performance and compliance. The penetration testing module can either be used as a no-code tool directly in the Dynamo AI platform or through the Dynamo AI python SDK solution. Currently, Dynamo AI offers 4 types of assessments: privacy, hallucination, compliance and performance. Each assessment type includes a variety of unique test types to assess vulnerabilities.

Attacks and Tests

Test Inventory

Test Category	Test Name	Test Detail
Privacy Test	PII Extraction	Tests whether these PII can be emitted from prompting the model naively and simulates an attacker with no knowledge of the training dataset
	PII Inference	Tests whether a model can re-fill PII into sentences from the fine-tuned dataset that we redacted PII from, assuming an attacker with knowledge of the concepts and potential PII in the dataset
	PII Reconstruction	Tests whether a model can re-fill PII into sentences from the fine-tuned dataset that we redacted PII from, assuming a partially-informed attacker
	Sequence Extraction	Tests whether the model can be prompted in a manner where it reveals training data verbatim as part of the responses
	Membership Inference	Determine whether specific data records can be inferred to be a part of the model's training dataset
Hallucination	NLI Consistency	Measures the logical consistency between an input text (or document) and a model-generated summary with an NLI model.
	UniEval Factuality	Measures the factual support between an input text (or document) and a model-generated summary with UniEval model.
RAG Hallucination	Faithfulness	Faithfulness represents how faithfulness model generated responses are to the set of retrieved documents.
	Retrieval Relevance	Retrieval relevance represents the relevance of the documents retrieved from the vector database using the embedding model for each query.
	Response Relevance	Response relevance measures how relevant the generated response is to the input query.
Compliance	Cybersecurity Compliance	Check whether model responses don't contain malicious code based on the MITRE framework
	Static Jailbreak tests framework	Tests whether the model is prone to jailbreaking attacks and if the model responses are compliant
	Bias/Toxicity tests framework	Tests whether the model can be forced to emit toxic or biased content
	Adaptive jailbreak attacks	Tests whether the model is prone to jailbreaks using algorithmically generated prompts
Performance	ROUGE	ROUGE is a metric designed to evaluate summary quality against a reference text by measuring the token-level overlap between the text generated by a model and the references.
	BERTScore	BERTScore computes semantic similarity between a reference text and model response, leveraging the pre-trained embeddings from the BERT model and is represented by a precision, recall, and F1 score.

Components

Architecture