AI Model Card

This model card provides a structured overview of the AI features and systems used at Intigriti.
Version: October 1st, 2025 (v1)

1. Introduction

This AI model card outlines the AI features available within the Intigriti platform, including their architectures, training data, performance metrics, intended uses, and limitations.

2. Architecture and Models

Models

Intigriti uses multiple AI models and technologies to power its platform:

  • Anthropic Claude (LLM - Amazon Bedrock)
    • Architecture: Closed-source transformer-based large language model, accessed via Amazon Bedrock in the EU region.
    • Usage: Provides contextual responses for text-based AI features.
  • Ember V1 (Sentence Transformer - Self-Hosted)
    • Architecture: Open-source bi-encoder transformer, deployed using the Sentence Transformers library.
    • Purpose: Converts text into dense vector embeddings.
    • Usage: Supports similarity matching, clustering, and downstream models.
  • XGBoost (Binary Classification - Self-Hosted)
    • Architecture: Open-source gradient-boosting framework optimized for structured data.
    • Objective: Binary classification tasks.

Deployment

Our AI models are deployed on the following infrastructure:

  • Ray Serve:
    • Objective: Provides scalable model serving for Sentence Transformers and XGBoost over gRPC.
    • Deployment: Deployed on Intigriti’s managed AWS EKS cluster in the EU region.
  • Amazon Bedrock:
    • Objective: Provides access to large language models (Claude) for NLP and reasoning tasks.
    • Deployment: Executes workloads entirely within the EU region to ensure compliance and minimize latency.
    • Data Handling: All requests are processed in real time without persistence. No data is stored or retained during inference.

Data Flows

The diagram below shows how data flows through our AI infrastructure, from user requests, through inference, to response delivery.

AI Data Flows
AI data flows within the Intigriti platform.

3. Training Data

Anthropic Claude (LLM)

  • Pre-trained by Anthropic on a broad dataset of text.
  • No fine-tuning performed with Intigriti data.

Ember V1 (Sentence Transformer)

  • Pre-trained on a large general-purpose text corpus for robust embeddings.
  • Adapted with internal domain-specific datasets (confidential).

In-house Models

In-house models are trained using metadata from customers with the AI Feature Flag enabled. This flag is enabled by default and can be disabled on request. Disabling it removes access to all AI-powered features.

The specific data used varies by feature. For details, see the Features and Models section.

4. Features and Models

This section describes the AI-powered features available within the Intigriti platform.

Submission Summarization

Submission Summarization is an AI feature that takes the contents of a single submission and generates a well-structured summary in a specific format. The feature is designed to help both customers and the triage team to quickly understand the context of a submission.

Architecture and Data

This model uses an LLM from the Anthropic Claude family with a specialized system and user prompt to generate a well-structured summary. Only the context of the current submission is provided to the model. The model is not trained on any customer data.

Performance Metrics

The model's performance is measured by human reviewers and standard metrics (ROUGE, BLUE) on internal test sets.

Intended Use

This feature is available to the internal triage team and customers with the AI feature flag enabled to streamline operations and decision making.

Submission Skill Labeling

Submission Skill Labeling is an AI feature that identifies the used skill for creating a submission.The feature is designed to facilitate required skill suggestions on an asset.

Architecture and Data

This model uses an LLM from the Anthropic Claude family with a specialized system and user prompt to predict the used skill. Only the context of the current submission is provided to the model. The model is not trained on any customer data.

Performance Metrics

The model's performance is measured by human reviewers and standard metrics (ROUGE, BLUE) on internal test sets.

Intended Use

This feature is available directly to the internal triage team. The data is also used to show required skill suggestions on company assets towards customers.

Submission Dupe Detection

Submission Dupe Detection is an AI feature that identifies potential duplicate submissions within the same program and company. The feature is designed to help the triage team spot duplicate submissions and reduce the amount of time it takes to review and act on these reports.

Architecture and Training Data

This model utilizes an in-house XGBoost binary classification model, trained on anonymized submission pairs that are labeled (duplicate vs. non-duplicate). It combines similarity scores derived from text embeddings with categorical data to identify if a submission is a potential duplicate.

The model is trained on the following data:

  • Title Similarity
  • Endpoint / Vulnerable Component Similarity
  • Asset Name Similarity
  • Company Asset Equality
  • Company Asset Type Equality
  • Type Equality
  • Type Category Equality
  • Severity Equality
  • Researcher Equality
  • Program Equality
  • Created At Delta (difference in creation time)

We do not train on data of companies that have the AI Feature Flag disabled.

Performance Metrics

The model's performance is measured by the following metrics:

  • Accuracy: 95% of correct duplicate vs. non-duplicate classifications.
  • Precision / Recall / F1-Score / Cross-Validation: Reflects the trade-off between false positives and false negatives.

The XGBoost model is re-trained every 6 months on resolved submissions, during which the accuracy and metrics are re-evaluated.

Intended Use

This feature is only available to the internal triage team to streamline operations and decision making.

Submission Similarity Detection

Submission Similarity Detection is an AI feature that identifies potential similar submissions within the same program and company. The feature is designed to help the triage team spot recurring resolved or rejected submissions.

Architecture and Training Data

This model utilizes an in-house XGBoost binary classification model, trained on anonymized submission pairs that are labeled (similar vs. non-similar). It combines similarity scores derived from text embeddings with categorical data to identify if a submission is a similar submission.

The model is trained on the following data:

  • Title Similarity
  • Endpoint / Vulnerable Component Similarity
  • Asset Name Similarity
  • Company Asset Equality
  • Company Asset Type Equality
  • Type Equality
  • Type Category Equality
  • Severity Equality

We do not train on data of companies that have the AI Feature Flag disabled.

Performance Metrics

The model's performance is measured by the following metrics:

  • Accuracy: 95% of correct similar vs. non-similar classifications.
  • Precision / Recall / F1-Score / Cross-Validation: Reflects the trade-off between false positives and false negatives.

Intended Use

This feature is only available to the internal triage team to streamline operations and decision making.

Submission Out-of-scope detection

Submission Out of Scope Detection is an AI feature that identifies if a given submission is out of scope, according to the program out of scope guidelines. The feature is designed to help the triage team spot out of scope submissions and reduce the amount of time it takes to review and act on these reports.

This model uses an LLM from the Anthropic Claude family with a specialized system and user prompt to detect if a submission is out of scope and which out of scope rule has been matched. Only the context of the current submission, program in scope and out of scope rules are provided to the model. The model is not trained or fine-tuned on any customer data.

Performance Metrics

The model's performance is measured by human reviewers and standard metrics (ROUGE, BLUE) on internal test sets.

Intended Use

This feature is only available to the internal triage team to streamline operations and decision making.

Submission Suggestions

Submission Suggestions is an AI feature generates on demand and automated title and endpoint suggestions, based upon the triage guidelines and data about the current submission. The feature is designed to help the triage team spot out of scope submissions and reduce the amount of time it takes to review and act on these reports.

Architecture and Data

This model uses an LLM from the Anthropic Claude family with a specialized system and user prompt to generate a well-structured title or endpoint. Only the context of the current submission, program in scope and out of scope rules are provided to the model. The model is not trained or fine-tuned on any customer data.

Performance Metrics

The model's performance is measured by human reviewers and standard metrics (ROUGE, BLUE) on internal test sets.

Intended Use

This feature is only available to the internal triage team to streamline operations and decision making.

Program Impact Generator

The program impact report generator is a tool designed to help Customer Success Managers lead impactful conversations with customers about how Intigriti contributes to their security posture.

Architecture and Data

This model uses an LLM from the Anthropic Claude family with a specialized system and user prompt to generate a well-structured report. Only the context of a set of selected submissions and program description are provided to the model. The model is not trained or fine-tuned on any customer data.

Performance Metrics

The model's performance is measured by human reviewers and standard metrics (ROUGE, BLUE) on internal test sets.

Intended Use

This feature is only available to the internal Customer Success team to help them understand technical submissions and lead conversations with customers.

5. Ethical Considerations

Data Privacy

  • All submission data is securely stored and processed under privacy guidelines.
  • Embeddings and aggregations minimize exposure of sensitive text.

Bias and Fairness

Models may inherit biases from training data. We monitor and mitigate bias through testing, human review, and performance metrics.

Human Oversight

AI features act as decision aids. Final responsibility remains with human reviewers.

6. Limitations

Large Language Models

  • May omit important details or introduce minor inaccuracies.
  • Human verification is recommended for high-stakes contexts.

Embedding Generalizability

Ember V1 embeddings may underperform on highly specialized or niche domains.

Resource Requirements

LLM inference via Amazon Bedrock is cost-efficient but requires monitoring and optimization.

7. Maintenance and Upgrades

Scheduled Retraining

In-house models are retrained every 6 months with updated labeled data.

Monitoring

We continuously track performance and alert on anomalies.

Versioning

Major updates to prompts, embeddings, or in-house models are documented for traceability.

8. Contact and Support

Technical Support

For questions, support, or incident reporting, please contact the support team.

Documentation

Further technical details and integration guides are available upon request.