How to Evaluate AI Security Vendors and Manages MCP Risks
Show notes
As organizations rush to adopt AI security tools, many discover that while these agents work well within their own silos (EDR, CSPM, SCA), they fail to communicate across boundaries, leaving security teams blind to the true blast radius of an incident. Additionally, the widespread use of the Model Context Protocol (MCP) by both engineers and non-technical staff is creating massive credential sprawl, often leaving unencrypted, over-privileged credentials sitting in plain text config files.
In this episode of the AI Security Podcast, Ashish speaks with Mohit Bansal, Senior Manager of Security Engineering at Webflow, about separating the signal from the noise when evaluating the explosive AI security vendor market. Mohit explains his framework for investing in structurally permanent solutions (like SCA) while waiting for the dust to settle on temporarily hard problems (like Agentic SOCs). The conversation dives deep into the realities of securing MCP calls, the need for open-source tools to redact secrets from agent memory, and why autonomous security still requires expert human knowledge to set blast radius boundaries.
Questions asked:
(00:00) Introduction to AI Vendor Evaluation and MCP Risks(02:50) How AI is Accelerating the Implementation Speed for Developers(05:30) Why Security Vendors Struggle to Identify the True Blast Radius(08:30) The Consolidation of the AI Security Vendor Market(12:30) Framework for Evaluating AI Vendors: Permanent vs. Temporary Problems(16:30) Managing Security Operations entirely through the Terminal via MCP(20:30) Why AI Will Not Eliminate the Need for Security Maintenance(23:30) Introducing "Context Cop": Webflow's Open Source Secret Redactor(27:30) How to Develop AI Autonomy Based on Engineering Seniority(32:30) The Danger of Blindly Trusting Non-Deterministic AI Output(35:30) Incident Response for AI: Visibility, Containment, and Tabletop Exercises(38:30) The Hidden Dangers of Plain Text Credentials in MCP Config Files(42:30) Advice for Security Teams with Limited Access to Frontier Models