Page 198 - Cyber Defense eMagazine RSAC Special Edition 2025

prompt violates a data rule for something we don't want going to an API like OpenAI or Anthropic, we
apply redaction at machine speed, without any user impact or user experience change."

It's also crucial to adopt a risk-based approach. Not all AI interactions carry the same level of risk. A
student asking about cafeteria hours doesn't need the same security controls as a faculty member
uploading research data.

What technical capabilities should security teams look for when securing AI implementations?

Security teams need several key capabilities when implementing AI security guardrails.

Content redaction in real-time is foundational – technology must identify and remove sensitive information
from prompts before they reach external AI models. An effective solution should recognize various data
types like student IDs, health information, financial details and other PII without introducing noticeable
delays.

Further, file sanitization capabilities become essential as more AI systems accept document uploads.
The security layer needs to scan and clean documents of sensitive metadata and content before
processing, preventing data leakage through attached files. GCE built their security approach around this
principle, seeking guardrails that could handle both conversational text and document uploads to their AI
platform.

Customizable policy controls allow security teams to define what constitutes sensitive information in their
specific context. This is crucial because organizations often have unique identifiers and data structures
that generic solutions might miss.

Performance cannot be compromised – any security layer should add minimal latency to maintain a
seamless user experience. When security introduces noticeable delays, user adoption inevitably suffers
and creates incentives for workarounds. For GCE, this performance requirement was essential as they
needed to deploy quickly without introducing latency that could affect student and faculty experiences.

Comprehensive logging and visibility round out these requirements, providing the audit trails needed for
compliance. Security teams need clear insights into what types of sensitive information are being caught
and redacted, without necessarily exposing the sensitive data itself.

What are the basics of a secure AI framework?

I recommend a four-phase approach:

Assessment: Begin by cataloging what AI systems are already in use, including shadow IT. Understand
what data flows through these systems and where the sensitive information resides.

Policy Development: Create clear policies about what types of information should never be shared with
AI systems and what safeguards need to be in place. Get buy-in from academic, administrative and
executive stakeholders.

198