Cloud Security Summit

Agentic AI is rapidly integrating into enterprise systems through the Model Context Protocol (MCP.) This introduces new and significant risks. Recent assessments reveal that 43% of tested servers are vulnerable to command injection, 82% use file operations susceptible to path traversal, and 33% contain critical vulnerabilities. Once a single server in a chain is compromised, the risk cascades to 72% of connected systems.

Sandeep Tripathi will examine real-world incidents including tool poisoning, rug pulls, indirect prompt injection, and supply chain remote code execution. He will outline a practical defense strategy: establish clear trust boundaries, categorize tools by risk level, require human-in-the-loop approval for state-changing actions, sandbox runtime environments, and enforce strict schema and parameter validation. Sandeep will also introduce an MCP gateway designed for risk scoring, authorization, validation, approvals, and rate and result limits, along with AWS and Azure guardrails supported by continuous monitoring.

Key takeaways:
- Quantify MCP risk with hard signals and real attack paths.
- Apply tiered policies: automatically approve read operations, and require human-in-the-loop approval for destructive or critical actions.
- Contain the blast radius by enforcing isolation and using least-privilege scopes with just-in-time tokens.
- Follow five essential rules: trust nothing by default, treat every server as a potential attack surface, authorize and audit each tool, minimize permissions, and require gating or reversal for material actions.