Bellini Capital CISO Insights: AI Security Governance & Patching Limitations

Brian Blakley, chief information security officer at Bellini Capital, addressed evolving dynamics between artificial intelligence advancements and cybersecurity practices.

The conversation explored a paradigm shift in Blakley’s perspective on AI security.

Two years ago, he believed AI would provide attackers and defenders with comparable advantages. This view changed after observing how AI integration in vulnerability management tools transformed discovery speeds. While threat actors can rapidly exploit newly identified flaws, defenders face complex workflows involving impact assessments, patch validation, compatibility testing, and coordination with business units.

Blakley cited his previous role at a fintech firm processing $50 billion annually in transactions, where system downtime for urgent patches was intentionally avoided to maintain revenue streams. This practice underscored a calculated tradeoff between exploitation risks and operational continuity.

Blakley proposed that the focus should shift from merely accelerating patches to minimizing exposure during remediation.

He outlined strategies such as network segmentation, isolation techniques, compensating controls, and expedited risk evaluations to align objectives between CIOs and CISOs. While CIOs prioritize project delivery and adoption metrics, CISOs measure success through risk mitigation. A unified performance metric, he suggested, could involve tracking the percentage of critical initiatives completed within financial and operational parameters while addressing material risks.

The discussion addressed mechanisms for testing AI model updates that could impact security outcomes.

Blakley noted that traditional change management practices assume stability when user-facing elements remain unchanged. However, AI models can alter internal reasoning without visible modifications to interfaces or version numbers. For example, a vulnerability prioritization system might reclassify critical flaws as low-priority without changes to the underlying environment or controls. This necessitates risk-based regression testing, where scenarios involving critical assets, known vulnerabilities, and business dependencies are continuously validated against established baselines.

Accountability for AI-driven incidents was another key topic.

Blakley described a framework where financial responsibility aligns with funding sources, operational accountability rests with teams managing incident response and regulatory compliance, and executive oversight is assigned to business owners approving AI deployments. He referenced a pilot program where agentic AI was used for outbound calls, requiring pre-defined guidelines to determine liability for errors. The framework emphasized that while AI can execute tasks faster, it cannot bypass human oversight for high-impact decisions.

Examples of AI tool restrictions were analyzed.

Blakley cited Walmart’s approach, where initial access blocks to ChatGPT allowed time for risk evaluation and policy development, resulting in a controlled implementation. Conversely, he highlighted cases where corporate restrictions led to shadow AI usage, as employees circumvented policies by using personal accounts. This underscored the importance of providing approved alternatives through governance committees that balance productivity needs with security requirements.

The discussion concluded with insights into AI governance structures.

Emphasizing the need for cross-functional collaboration between business units, technology leaders, and security teams to address evolving threats while maintaining operational efficiency.