Topic
Insights and notes grouped by this theme.
The model is not the hard part. The hard part is everything around the model.
The modern data stack is powerful, but operationally fragile. AI makes the weak points obvious: access control, contracts, lineage, observability, cost, and security.
Agentic AI is powerful because it can act. That is also why it needs real controls.
AI agents are moving from demos into production workflows. Once they call APIs, write records, and trigger business processes, they become part of the control plane.
What ServiceNow's AI agent incidents reveal about enterprise workflow integrity.
Prompt injection becomes materially more dangerous when AI agents can act. Once language is connected to workflows, text becomes a control input.
Generative AI risk does not fit neatly inside traditional security controls.
CISOs need a practical AI governance model that connects acceptable use, data controls, model risk, third-party exposure, monitoring, and incident response.
The uncomfortable part is simple: language has become an attack surface.
Prompt injection works because large language models do not reliably separate instructions from data. Once a model can use tools, retrieve documents, or act on systems, that weakness becomes an enterprise security problem.
Model security starts upstream. Bad data governance becomes AI security risk.
AI security begins before model deployment. Poorly governed, unclassified, or unmonitored data can create vulnerabilities that cascade through the entire AI lifecycle.
