Panel Discussion: Proof in Production: Customer Lessons from Confidential AI

### Overview This panel discussion, moderated by Aaron, explores the operational and security challenges of deploying AI and autonomous agents across highly regulated industries, including financial services, fraud prevention, digital healthcare, and enterprise software. Industry leaders discuss why traditional reliance on simple trust, narrative assurances, and post-facto compliance is no longer sufficient. The discussion centers on the shift toward confidential computing, runtime verification, and cryptographic proof to enable secure, compliant data processing across distributed environments.

### Key Topics - **Moving Beyond Trust to Verifiable Proof**: Transitioning from passive assurances, data anonymization, and legal attestations to concrete technical artifacts, verifiable integrity, and runtime attestation. - **Regulatory Pressures & Runtime Evidence**: Regulators adopting AI to conduct exhaustive audits and generate extensive compliance questionnaires, alongside emerging regulatory mandates (such as EU AI Act runtime proof requirements) that demand continuous operational transparency. - **Agentic Security & Three-Tier Governance**: Managing agent risks across preventative, runtime, and detective layers, including fine-grained tool controls, real-time risk monitoring (such as ServiceNow's AI Control Tower / AICT), and immutable, cryptographically signed audit trails. - **Cross-Enterprise Data Challenges**: Overcoming regulatory barriers in fraud detection and healthcare where sensitive data (PII, biometrics, PHI, multimodal surgical video/images) cannot be simply shared or centralized without verifiable guarantees and confidential execution environments.

### Technical Details - **Confidential Computing & Attestation**: Panelists highlight confidential computing as a mechanism to produce verifiable artifacts that prove exactly what code and models ran, on which infrastructure, and against what data sets without exposing sensitive data. - **Fine-Grained Tool Enforcement & Confidential MCP**: Collaboration around confidential Model Context Protocol (MCP) implementations introduces preventative pre-deployment scanning down to fine-grained tool action levels (e.g., restricting tools strictly to read operations versus write operations), cryptographically enforced via confidential environments. - **Multi-Stage Distributed Pipeline Tracking**: In healthcare surgery platforms, data moves from localized provider systems through regional clouds with sovereign policy constraints before reaching analytical pipelines. Verifiable computing enables tracking what happened, what changed, and authorization status across every discrete processing step. - **Three-Tier Architectural Governance**: ServiceNow structures agent governance across: 1. *Preventative*: MCP scans, MCP server audits, and granular tool-level capability controls. 2. *Runtime*: Continuous monitoring via an AI Control Tower (AICT) tracking active agents, runtime behaviors, and risk scores. 3. *Detective*: Post-execution immutable, cryptographically signed audit logs exportable for compliance.

### Takeaways - **Establish Structured Governance Early**: Organizations must design robust governance frameworks covering performance, security, ethics, and legality before scaling AI and agent deployments. - **Prioritize Data Classification & Cross-Functional Alignment**: Success requires strict data classification alongside close collaboration across engineering, operations, compliance, CI/CD, and legal teams. - **Deploy Internal Adversarial Tiger Teams**: Companies should form small internal teams tasked with attacking their own non-production pipelines, models, and agents to surface vulnerabilities before external auditors or adversaries do. - **Iterate in Low-Risk Environments**: Experiment with AI and build out governance processes on internal, non-customer-facing systems before applying them to mission-critical or highly sensitive customer data.