What is Prisma Cloud Palo Alto?
A complete cloud-native application protection platform (CNAPP), Prisma Cloud Palo Alto offers security and compliance coverage for workloads, apps, and infrastructure in public, private, hybrid, and multi-cloud settings. It aids businesses in lowering risk, enhancing compliance and security posture, and improving Dev-Sec collaboration.
By offering a route from prototype to production, Google Cloud enables companies to quicken their generative AI innovation cycle. Global cybersecurity pioneer Palo Alto Networks teamed up with Google Cloud to create a cutting-edge security posture control system that can assist users through remediation procedures, offer deep insights into risk with a few clicks, and deliver sophisticated “how-to” answers on demand.
Palo Alto Networks has the perfect platform to develop and implement new AI-powered solutions by using managed Retrieval Augmented Generation (RAG) services like Google Cloud’s Vertex AI Search and advanced AI services like Google’s Gemini models.
The outcome was the Palo Alto Prisma Cloud gen AI product, Prisma Cloud Co-pilot. It makes cloud security management easier by offering a user-friendly interface driven by AI to assist identify and reduce threats.
Technical challenges and surprises
Beginning in 2023, the Prisma Cloud Palo Alto Co-pilot project was started in October 2024. Palo Alto Networks saw Google’s AI models advance quickly during this period, from Text Bison (PaLM) to Gemini Flash 1.5. Every iteration introduced new capabilities due to the high speed of invention, therefore a development method that could swiftly adjust to the changing environment was required.
In order to successfully negotiate the ever-changing terrain of developing AI models, Palo Alto Networks put in place reliable procedures that were crucial to their success:
Prompt engineering and management
Palo Alto Networks created a varied prompt library to provide a variety of replies and using Vertex AI to assist in managing prompt templates. The Palo Alto Networks and Google Cloud team methodically developed and updated prompts for every submodule in order to thoroughly evaluate each new model’s capabilities, constraints, and performance across a range of activities. Furthermore, the time-consuming trial-and-error procedure of prompt engineering was streamlined using Vertex AI’s Prompt Optimiser.
Intent recognition
Palo Alto Networks created an intent recognition module based on the Gemini Flash 1.5 concept that effectively directed user enquiries to the appropriate co-pilot component. This method gave consumers a lot of features in a single, lightweight user interface.
Input guardrails
As a first line of defence against unexpected, malicious, or just plain wrong questions that might jeopardise the chatbot’s functionality and user experience, Palo Alto Networks developed guardrails. By limiting chatbot usage to its intended scope and preventing known prompt injection threats, including evading system commands, these guardrails preserve the chatbot’s intended functionality.
In order to avoid accidental usage, guardrails were developed to determine if user searches are limited to answers falling within the predetermined realm of general cloud security, risks, and vulnerabilities. The chatbot did not respond to any topics that fell outside of this range. Furthermore, requests for general-purpose code creation were not answered because the chatbot was created for the purpose of generating proprietary code for Palo Alto Networks systems to query internal systems.
Evaluation dataset curation
The basis for precisely and rapidly evaluating the performance of various AI models is a strong and representative assessment dataset. By selecting high-quality assessment data and continuously updating it with representative questions and expert-validated responses, the Palo Alto Networks team made sure the data remained current. Palo Alto Networks subject matter experts were consulted directly to verify the evaluation dataset’s quality and dependability.
Automated evaluation
Palo Alto Systems and Google Cloud automate evaluating with Vertex AI’s gen AI evaluation service. With the use of this pipeline, Palo Alto Networks was able to thoroughly scale their evaluation of several generation AI models and benchmark them using unique evaluation criteria, all the while concentrating on important performance metrics like response consistency, accuracy, and latency.
Human evaluator training and red teaming
Palo Alto Networks made the investment to teach their human review staff to recognise and examine certain loss patterns and offer thorough responses on a wide range of unique rubrics. This enabled them to identify areas in which a model’s reaction was insufficient and offer perceptive feedback on model performance, which subsequently directed the selection and improvement of models.
Load testing
For optimised real-time performance, Palo Alto Networks tests current gen algorithms per Gemini model QPM and latency. They used provided throughput to model user traffic scenarios and determine the best balance between scalability and responsiveness, ensuring a seamless user experience even during periods of high usage.
Operational and business challenges
Putting modern AI into practice can provide difficult problems for a variety of departments, including information security, legal, and compliance. New measures are also needed to assess the return on investment of emerging AI systems. Palo Alto Networks used the following methods and procedures to deal with these issues:
Data residency and regional ML processing
It gave regional machine learning processing top priority since many Palo Alto Networks clients require a regional approach for ML processing capabilities. This will assist customers comply with data residency requirements and, if relevant, regional legislation.
Customers had the option to have their data processed in the United States prior to receiving access to the Prisma Cloud Co-pilot in cases where Google did not provide an AI data center that matched Prisma Cloud data center locations. To assist protect sensitive data and preserve user privacy, they put in place stringent data governance procedures and made use of Google Cloud’s secure architecture.
Deciding KPIs and measuring success for gen AI apps
Because gen AI applications are dynamic and complex, a unique set of metrics is required to fully assess their effectiveness and capture their unique features. No single set of measurements is appropriate for every use scenario. Technical and business indicators were used by the Prisma Cloud AI Co-pilot team to gauge the system’s performance.
- In order to improve the accuracy of rapid replies and give consumers source information, technical measures like recall were used to gauge how completely the system retrieves pertinent URLs while responding to enquiries from documents.
- Measures of customer experience, including helpfulness, depended on telemetry data analysis and explicit input. Deeper understanding of the user experience was made possible by this, which raised output and reduced expenses.
Collaborating with security and legal teams
In order to identify risks and establish safeguards for issues such as information security requirements, bias elimination in the dataset, appropriate tool functionality, and data usage in compliance with applicable law and contractual obligations, Palo Alto Networks included legal, information security, and other crucial stakeholders early in the process.
Businesses must place a high priority on transparent communication on data usage, storage, and protection in light of consumer concerns. Palo Alto Networks was able to gain the trust of its customers and make sure they understood how and when their data was being used by working with the legal and information security teams early on to provide openness in marketing and product communications.
Ready to get started with Vertex AI ?
Generative AI has a bright future, and businesses may realize its full potential with proper planning and implementation. Use Vertex AI’s useful pilots to investigate your company’s AI requirements, and get professional advice from Google Cloud Consulting.