GenAI & SDLC (Software Development Lifecycle)
As a cloud-native SaaS provider, our AI platform follows a risk-based governance approach and, as of the time of writing, our AI capabilities are internally assessed as low risk based on their intended use and system design, subject to ongoing review.
We distinguish our models between shared/models and customer-specific. Shared customer models consist of large scale, fine-tuned models used for general language understanding, feedback experience analysis and reasoning capabilities. Models act as a processing engine, not a storage point, therefore, client data is never retained beyond the contractual limitation period.
Capabilities such as Smart Response, Root Cause Assist, Intelligent Summaries and Coaching Intelligence are all cross-customer models. Customer-specific models are dedicated components allowing the AI to understand your unique business context and use case. These assets are logically isolated within your tenant boundary, ensuring privacy and making it inaccessible to any other Medallia client.
Themes with GenAI is the sole GenAI capability that inhibits customer-specific models per tenant. Together, this approach reduces operational complexity, limits unintended and uncontrolled data flows, and provides enterprises with a secure and transparent foundation for responsible AI adoption.
Security By Design at Medallia
Medallia in addition to our expanded secure SDLC for generative AI, applies secure system prompt engineering practices aligned with industry standard security frameworks such as OWASP for Large Language Model Applications and relevant SANS secure AI guidance at this time.
These controls include strict PII minimization (in adherence to GDPR Article 5), structured input validation, and robust code injection protections through escaping and sanitization mechanisms. Human-in-the-loop (HITL) controls are incorporated where appropriate to provide oversight, exception handling, and controlled decision points for higher-risk interactions.
To mitigate common GenAI attack vectors, we extend our existing SaaS security controls to address risks such as LLM model manipulation and prompt-based exploitation, human process weaknesses, overly permissive GenAI feature configurations. These protections are informed by OWASP and SANS threat models and are continuously reviewed as new GenAI-specific risks emerge.
We further strengthen our security posture through ongoing GenAI internal and external red teaming and product security reviews , to validate model behavior, control effectiveness, and operational resilience against realistic attack techniques. In parallel, our vendor risk and supply chain management program applies defined risk mitigation criteria—aligned with OWASP and SANS best practices—when selecting and managing third-party technologies, ensuring that all supporting components meet our security, compliance, and operational standards.
Medallia also protects its compatible solution offerings at the presentation layer, deploying runtime application self-protection (RASP) controls that continuously monitor and inspect common attack vectors such as XSS, Injection attempts as well as now our GenAI interactions in real time. These controls are evolving to detect and mitigate common Generative AI attack vectors, including prompt manipulation, injection attempts, and anomalous input or output patterns. By enforcing protective actions at runtime, RASP enables immediate response to emerging GenAI vulnerabilities without disrupting application availability or user experience.
