AI Accelerates Vulnerability Discovery, But Patching Remains a Bottleneck, Says Bellini Capital CISO
Bellini Capital CISO Brian Blakley argues that while AI dramatically speeds up vulnerability discovery, the traditional patching process remains a significant bottleneck, necessitating new strategies to reduce exposure during remediation.
In an interview with Help Net Security, Brian Blakley, CISO at Bellini Capital, shared his evolving perspective on the impact of artificial intelligence on cybersecurity. Blakley initially believed AI would offer comparable advantages to both attackers and defenders. However, firsthand experience managing a vulnerability management platform has led him to abandon this view. He now observes that AI is significantly accelerating the discovery of vulnerabilities, a trend that benefits both offensive and defensive security operations.
Blakley highlights a critical asymmetry: while AI empowers attackers to move from discovering a flaw to deploying a working exploit with unprecedented speed and minimal friction, defenders face a protracted remediation process. This process involves assessing business impact, validating patches, ensuring compatibility, obtaining change approval, and scheduling maintenance windows. For critical systems, such as payment platforms processing billions of dollars, taking systems offline for patching is often a deliberate, risk-averse decision, balancing the threat of exploitation against the certainty of revenue interruption.
Consequently, Blakley posits that the common security mantra of "patch faster" is insufficient. The more pertinent challenge, he argues, is to "shrink exposure while remediation is pending." This involves leveraging techniques like segmentation, isolation, and compensating controls, alongside faster business risk assessment and decision-making. This approach aligns the objectives of both CIOs and CISOs, focusing on maintaining critical services while rapidly closing exploitable windows.
Addressing the perennial challenge of misaligned incentives between IT and security leadership, Blakley advocates for shared metrics that benefit the entire organization. He criticulates "split scorecards" where a CIO might meet delivery targets by introducing risk, or a CISO might reduce exposure by delaying projects without clear business justification. A more effective model, observed at a large insurance-adjacent company, involves funding AI initiatives through small proof-of-value projects with a clear objective: grow revenue while keeping labor costs flat. The AI and Data Governance Committee, comprising business, operations, IT, and security leaders, vets initiatives against this outcome, using a shared metric of critical initiatives delivered within financial and operational constraints, with material risks explicitly accepted by business owners.
Blakley also emphasizes the need for robust testing of AI model updates, which can introduce subtle yet significant behavioral changes without altering interfaces or version numbers. For instance, a vulnerability prioritization model could shift its risk assessment, downgrading a critical vulnerability to low priority without any apparent change to the system or its controls. To counter this, Bellini Capital employs risk-based regression testing, comparing outputs against validated baselines for shifts in risk interpretation, recommended actions, and data handling. While they haven't experienced a security incident directly caused by a silent model update, Blakley cautions that this may only indicate a failure to measure the right aspects of AI behavior.
Finally, Blakley addresses the critical issue of accountability for AI agent actions. He asserts that while organizations can delegate authority to AI agents, they cannot delegate accountability. When an AI agent with delegated permissions causes damage, such as deleting records or mishandling customer data, the responsibility ultimately rests with human leadership. The cleanup costs and board reporting fall under existing executive purview, underscoring the need for clear governance frameworks and defined lines of responsibility before deploying autonomous AI agents.