Skip to content

AI Detection & Response

AI Detection & Response (AIDR) provides flexibility in detecting or blocking content being sent to an LLM or returned to the user.

The HiddenLayer AI Detection & Response is a real-time input and output monitor for hosted or custom LLMs. It detects malicious input prompts and/or undesired output as they are sent to and returned from an LLM, and can (when configured appropriately) block content from being sent to the LLM or returned to the user. It has different modes of operation which can be flexibly employed, depending on the architecture already in place and the desired level of integration.

HiddenLayer’s target operating model is designed to provide maximum flexibility, security, and operational independence for our customers. Our software is available as container images, allowing for seamless deployment, scaling, and integration into existing customer architecture. We provide pre-packaged, production-ready container images, which the customer deploys, configures, and operates independently within its own cloud or on-premises Kubernetes infrastructure. This makes deployment and integration into an existing containerized infrastructure straightforward for DevOps teams.

Additionally, since the software is provided as a containerized application accessible via API, it is not only available on the cloud but can also be deployed as a locally running container on a developer workstation for educational, evaluative, or development purposes. This means that application developers or data scientists, who may not have a background in cloud architecture or the necessary access to resources, can follow our simple, step-by-step user guides to spin up a locally running instance of the software with all of the same API-accessible capabilities of the cloud deployment. They can then use the locally running instance to test out integration code, develop application connectors or data visualization tools for result sets, and to run small-scale pre-deployment test sets against the software to validate performance, all without the overhead of a full cloud deployment or the need for additional resources, and without interrupting production workflows until they are ready to deploy their changes to production. Because the cloud containers and the local containers are built using the same images and source code, adapting the code for use between development and production is as simple as changing the endpoint used to access the HiddenLayer instance.

HiddenLayer Support

For help with the HiddenLayer platform, email Support at support@hiddenlayer.com.

Key Benefits

  • Automated: Leverage automated processes to detect and respond to AI model breach attempts, providing a proactive defense mechanism.
  • Scalable: Get clear reporting on detected threats, empowering security teams with insights into adversarial behavior. Identify and report on various adversarial activities such as model theft, reconnaissance, evasion, misclassification, and other potential threats.
  • Unobtrusive: Detections are made via our platform without requiring any access to models and the data that powers them.

Key Capabilities

  • Prompt Injection: Ensure inputs to your LLM do not cause unintended consequences.
  • Data Leakage: Ensure LLM outputs do not expose backend systems risking privilege escalation or remove code execution.
  • MITRE ATLAS Integration: MLDR maps to 64+ Adversarial AI attack tactics & techniques.
  • Inference Attacks: Protects against real-time model Inference Attacks.
  • Protects against Model Tampering: Know where the model is weak and when the model has been tampered with.
  • Protects against Prompt Injection/Model Injection: Protect the model from its inputs or outputs being deliberately changed.
  • Protects against Model Extraction/Theft: Stop reconnaissance attempts through inference attacks which could result in your model intellectual property being stolen.
  • Combined Methods for Detections: Uses a combination of Supervised Learning, Unsupervised Learning, Dynamic/Behavioral Analysis, and Static Analysis to deliver detections for a library of adversarial machine learning attacks.

End User License Agreement

Usage of this Docker Image is subject to End-User License Agreement (EULA). Before accessing or utilizing the Docker Image, please carefully review and agree to the terms and conditions outlined in the EULA provided by HiddenLayer. The EULA governs the rights, limitations, and obligations associated with the use of the Docker Image. By using the Docker Image, you indicate your acceptance of the EULA and your commitment to adhere to its provisions. If you do not agree with the terms and conditions set forth in the EULA, refrain from using the Docker Image.