# Snyk's Manoj Nair: Skills and MCP Are the New npm

## Executive summary

The adoption of AI coding agents is rapidly transforming the software development lifecycle, moving beyond simple co-piloting to fully autonomous 'software factory' capabilities. While this acceleration promises massive productivity gains, it introduces significant security challenges. Experts emphasize that securing the process (the factory) is now as critical as securing the product itself. Key risks include insecure agent output, the supply chain vulnerability of new components (like Skills and MCP servers), and unpredictable agent behavior. Overcoming these hurdles requires implementing advanced controls such as sandboxing, granular identity management, and shifting organizational risk acceptance to enable automated remediation and merging.

## Key takeaways

- The Shift to Agentic Development: The industry is moving from 'using AI' (Co-pilot era) to 'building with agents' (agentic development), where agents perform complex tasks autonomously, potentially running without constant human oversight [00:11:18].
- The Three Core Agentic Risks: The primary security concerns are threefold: insecure code output, the vulnerability of the new supply chain (e.g., Skills and MCP servers being the 'new npm'), and unpredictable agent behavior (e.g., an agent deleting a production database) [00:11:18].
- Security Must Secure the Factory: Security efforts must evolve from merely securing the final product to securing the entire 'software factory'—the process, the agents, and the workflows themselves [00:27:56].
- The Need for Organizational Risk Acceptance: Accelerating agent adoption requires a top-down cultural shift where organizations are willing to accept a calculated level of operational risk in exchange for fixing massive amounts of technical debt and improving speed [00:22:54].

## Technical details

- Agentic Supply Chain Risk: The supply chain risk has expanded beyond traditional dependencies. New components like Skills and MCP servers are viewed as the 'new NPM,' requiring security context and validation for every component used by an agent [00:11:18].
- Securing the Agent Loop: Best practices for running agents include using secure sandboxes (e.g., Docker), enforcing separate agent identities, and implementing dynamic/ephemeral authorization to limit the agent's blast radius [00:27:56].
- Automated Remediation: Specialized remediation agents can address technical debt and backlogs by using security context and global context (e.g., analyzing millions of repos) to suggest safe upgrade paths and fix vulnerabilities in a batch mode, achieving high success rates (e.g., 93% marked success rate) [00:33:00].
- Future Agent Controls: The next frontier in security requires solving for agent *intent* and *behavior* on the fly, rather than just blocking known threats. This involves understanding the business logic and the true purpose of the agent's request to grant precise, temporary permissions [00:40:18].

## Practical implications

- Shift security focus from manual AppSec gates to integrating security context directly into the agentic development loop.
- Implement granular controls (sandboxes, identity, ephemeral authorization) to minimize the blast radius of autonomous agents.
- Prioritize building automated remediation capabilities to tackle accumulated technical debt and security backlogs.
- Establish clear organizational policies that define and accept shared risk for the benefits of AI-driven speed and productivity.

## Topics

AI, Agentic Coding, Software Factory, Security, Technical Debt, CI/CD, Snyk, Skills and MCP servers, Tessl

Source: https://www.youtube.com/watch?v=URFAQgm86UA