Bellini Capital’s CISO says AI security needs more than faster patching

AI can speed up finding vulnerabilities, but fixing them still means weighing business impact, testing patches and arranging downtime. Brian Blakley, Bellini Capital’s CISO, argues that security teams should also reduce exposure while a fix is pending, and that organisations need clearer rules for AI tools and agents.

Watch Desk analysis

What happened

In an interview with Help Net Security, Blakley describes a security problem with an awkwardly human bottleneck: a vulnerability may be found quickly, but applying a patch can require compatibility checks, approvals and a maintenance window. He recommends shrinking the exploitable window with measures such as segmentation, isolation and compensating controls while remediation is pending.

He also describes risk-based regression tests for changes to underlying AI models. His team compares outputs against validated baselines, including changes to risk interpretation, recommended actions, data handling and permissions. He says providers do not always allow models to be pinned, and cautions that not observing a problem does not prove a silent update caused none.

On governance, Blakley argues that organisations should decide in advance who funds incident clean-up, who handles operational recovery and which business owner is accountable for an agent’s actions. He also warns that banning AI tools without offering an approved alternative can push use out of sight.

Why it matters

This is a practical account of AI security as an operational problem, not simply a race to patch faster. The advice is to limit exposure while fixes are in progress, test for consequential model changes and set responsibility before an agent acts. Blakley is describing his own experience and recommendations, not reporting a measured industry-wide result.

Our read

The strongest point is that security controls have to work with the way organisations actually operate. A ban can buy time, but it is not a strategy if staff still need the tool and find another route to it. Likewise, a model’s unchanged version number is not much comfort if its behaviour has shifted underneath.

What to watch

• Whether teams adopt regression tests for changes in model behaviour, not just software settings.
• Whether AI governance assigns clear business accountability for agents’ decisions.
• Whether restrictions come with approved tools that meet the work people need to do. Blakley’s interview offers a useful practitioner's perspective on AI security governance; the advice is worth considering, not a substitute for evidence about any particular organisation’s controls.

Discussion spark: When a company restricts an AI tool, should it be responsible for providing a safe alternative, or is that a separate business decision?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.