Skip to content

AI Generated Case Summaries

This document gives the key aspects of our GenAI implementation. It tells you how we handle data, and it gives information about privacy, usage, and the opt-out option.

AI Case Summaries are one part of the broader AI capability of SenseOn. Horus, the AI orchestrator of SenseOn, coordinates this capability. For information about how these parts work together, refer to SenseOn AI Architecture.


Accessing AI Summaries via API

You can use the SenseOn Connect API to get AI case summaries programmatically:

  • POST /v1/cases - Includes the AI summaries that already exist on cases
  • POST /v1/cases/{case_id}/summary - Generates a new LLM summary for a specific case, or generates the summary again
  • GET /v1/cases/{case_id} - Retrieves case details. The details include summaries only if the summaries already exist

For more information about the SenseOn Connect API, refer to the API Reference documentation.

AI Generated Case Summaries: How It Works

Google Vertex AI

SenseOn AI Case Summaries use Gemini 1.5 Pro, hosted on Google Vertex AI (EU region). SenseOn also uses Langfuse to monitor, analyse, and improve the quality and efficiency of our features that use generative AI. This gives you a better experience.

Data Usage and Retention

Google will not use your data to train its models. Google is a data sub-processor for SenseOn. Our agreement with Google does not let Google use any of your data to train its public models. For more details, refer to the Google documentation.

Data Processing: We chose to use the EU hosting region of Google Cloud. Thus, Google Cloud processes your data only in the EU.

Cross-border Data Transfers: Google Cloud services, which include Vertex AI, obey international data transfer standards. Examples are the EU-U.S. Data Privacy Framework and the Standard Contractual Clauses (SCCs). These standards make sure that cross-border transfers are lawful when they are necessary.

Customer data shared with Vertex AI

When SenseOn generates AI Case Summaries, it can send this data to the LLM:

  1. Case Information:
    • Case ID
    • Case Score
    • Case Description
    • Case Status
    • Timestamps
    • Observation details
    • MITRE techniques
  2. Device Information:
    • Host name
    • Host Description
    • User
    • Endpoint ID
    • IP Address (including country and city)
  3. Device Process Information:
    • Process name
    • Process PID
    • Process executable path
  4. Observation evidence (associated with the case):
    • SenseOn can also return a subset of the telemetry that is relevant to the observation to the LLM, for summarisation. This subset is the evidence of the observation.
    • SenseOn can return the telemetry in the “endpoint process” and “network process” tables to the LLM. This depends on the observation type.
    • Refer to the Table references in HuntLab for “network process” and “endpoint process”. They give the complete list of telemetry that is available to the LLM.
  5. Customer details:
    • Anonymised customer identifier (SenseOn issues a codename for each customer).

Model Input and Response Storage Retention * The cache keeps the inputs and outputs that the model processes for SenseOn Case Summarisation for up to 24 hours. * The cache lets Gemini models respond more quickly to later prompts. An example is when you refresh the summary to include new observations.

Langfuse

For more information, refer to the Langfuse data privacy and security documentation.

Customer data shared with Langfuse Langfuse stores a copy of the input and response of the model. Thus, Langfuse gets the same data that Vertex gets.

When an end-user gives feedback on a summary that SenseOn generated, Langfuse stores the response and the internal user ID of that user.

Data Processing: We chose to use the EU hosting region of Langfuse. Thus, Langfuse stores and processes all data for us in the European Union. This makes sure that the processing complies with EU data protection regulations.

Langfuse Data Storage Retention * Langfuse keeps end-user feedback indefinitely. It also keeps the model inputs and outputs for SenseOn Case Summarisation indefinitely.

Potential Changes in Data Processing

When we add new AI capabilities, the types and quantity of data that the LLM processes can change:

  1. Data Types: New features can make it necessary for the LLM to process more types of data. For example, future improvements can include the analysis of new forms of telemetry or threat intelligence data.
  2. Data Volume: We send data to the LLM for processing. The quantity of this data can increase or decrease when we optimise our AI integration.
  3. Processing Frequency: The frequency of AI data processing can change for each feature. It can become more real-time or more batch-oriented for specific use cases.
  4. Model & Provider Updates: We can update or change the LLM models that we use. This can change how the models process and analyse data.

Our Commitment to Transparency and Privacy

When we develop our AI capabilities, we will continue to:

  • Tell you about significant changes in AI features and data processing.
  • Notify you if we add a new data sub-processor to the list of agreed sub-processors. An example is a new model provider.
  • Keep strict data privacy and security standards, independently of AI improvements.
  • Give clear documentation about the data that AI models process.
  • Give granular control over the use of AI features, with opt-out options for specific AI functions.

AI Training and Improvement

  • You can give feedback on the AI generated case summary in the SenseOn UI. To do this, select thumbs up/down for each summary. When you select a thumbs down rating, you can also give free text feedback.
  • Our development team uses this feedback to improve and update the quality of the AI output
  • We make all improvements or updates to the AI capabilities through controlled processes

AI Generated Case Summaries request for review

Ethical Use and Transparency

  • We are committed to the responsible and ethical use of AI technology.
  • We designed our AI to help and augment human decisions, not to replace them.
  • We are transparent about the capabilities and limitations of the AI.

Opt-Out Option

We understand that some users or organisations possibly prefer not to use AI-assisted features. For this reason, SenseOn gives an opt-out option. We will respect your decision, and it will not affect your current services with us.

  • To opt out of AI-generated case summaries and other AI-assisted features, contact your SenseOn account manager or our support team.
  • If you opt out, Case Summaries will not be available in your SenseOn platform.
  • Our standard, non-AI systems will continue to process your case data and other information. This keeps the threat detection and response capabilities.
  • You can opt in again at any time. To do this, contact us.

If you opt out of AI features, this can affect some functions and the timeliness of threat response. Our support team can give more details about how an opt-out can affect your specific use case.

Limitations of AI Case Summary

We designed our AI Case Summaries to improve your experience with SenseOn. But it is important to understand their limitations. Our AI Case Summaries have some constraints, as all AI systems that use large language models do:

Potential Inaccuracies and Hallucinations

  • The AI can occasionally give inaccurate or inconsistent information.
  • It can misinterpret context or nuances in complex cybersecurity scenarios.
  • The AI can sometimes generate "hallucinations". A hallucination is information that seems correct, but that the AI fabricated completely.

Our development team continuously monitors the quality of the model output and makes adjustments. This minimises the likelihood of inaccuracies and hallucinations.

Data Freshness

The summary does not update automatically when the case gets new observations. It also does not update after actions occur (for example, a device is isolated or a case is closed). Thus, it is possible that the information in the summary is not up to date. You can generate a new summary at any time to get a summary with the case information.

Lack of Real-Time Knowledge

  • The knowledge of the AI comes from its training data. The training data has a cutoff date.
  • It is possible that the AI does not have information about the latest threats or vulnerabilities.

Contextual Misunderstandings

The AI can misunderstand the full context of a situation, especially in complex or unusual cybersecurity scenarios.

Mitigating These Limitations

To decrease the effect of these limitations, we recommend that you:

  • Always do a check of critical information that the AI gives.
  • Use AI Case Summaries as a tool to support human expertise, not as a replacement for it.
  • Regularly give feedback to help us improve the system.
  • Get information about the latest updates to the AI capabilities of SenseOn.

Contact our support team for more detailed information about how we handle data or about the opt-out process. Contact them also if you have concerns. As an alternative, contact dan.harrison@senseon.io (Product Manager) directly.