Version - 2
Date: 29th July, 2024
Improve performance of the following detectors by refactoring code
LLM Guardrails Detectors:
Prompt injection (Input scanner only)
Toxicity (Input and Output scanner)
gibberish (Input and Output scanner)
Language (Input and Output scanner)
Language Same (output scanner only)
This refactoring has enhanced the efficiency and accuracy of the scanners, resulting in an approximate 70% performance improvement in handling and processing input and output data.
Allow-Alert for LLM detector(s)
Implement a new feature to allow alert for LLM detectors. This feature permits traffic flow while generating alerts afterwards. It ensures that traffic is not blocked, providing seamless operation while still monitoring and reporting on the detected issues.
Running Parallel LLM detectors in LLM
LLM Parallel Detector: All LLM detectors now operate in parallel mode, significantly enhancing performance when all detectors are selected within a rule. This parallel processing capability ensures faster and more efficient detection.
Running Parallel LLM detectors in extension
Extension Parallel Detector: When detectors are chosen in the extension policy, all LLM detectors now operate in parallel mode. This significantly enhances performance, ensuring faster and more efficient detection of potential issues.
Implementation of Granular policy in LLM
LLM Granular Policy: A new granular policy design has been introduced in LLM. This implementation allows for more precise and detailed policy configurations, enhancing the ability to tailor and manage LLM operations to meet specific organizational requirements.
Implementation of nudity detectors
Extension Nudity Detector: When uploading an image containing nudity to AI sites, the nudity detectors will identify and flag the content accordingly. This implementation ensures that images with nudity are promptly detected and addressed.
Timeout Feature
Nucleuz Timeout Feature: The Nucleuz timeout feature is applicable when the server response takes too long to evaluate content by the Nucleuz engine. If the server response takes more than 2 seconds, the user will be allowed access, and any sensitive content detected will generate an alert asynchronously. A timeout message will be added to the alert.
Customize Detector List
LLM Guardrails: A new feature has been implemented that allows users to customize the detector in a novel way. This enhancement provides users with greater flexibility and control, enabling them to tailor the detection process to better meet their specific needs and requirements.
New Fields in Alerts, Events
Alerts/Events Details Page: Four new fields, "File Size", "Policy Violation", "Full Url" and "User Agent" have been added to the Alert and Event details page for Extension.
Audit Log features
When performing any action in the UI, an audit log is automatically created. This feature ensures that all activities are recorded for tracking and review purposes.
Last updated