MCP Security: Why Your AI Assistant Is an Insider Threat (with Liran Tal, Snyk) скачать в хорошем качестве

MCP Security: Why Your AI Assistant Is an Insider Threat (with Liran Tal, Snyk)

In this episode of The Context, we’re joined by Liran Tal (Snyk) to dig into MCP security and why “easy to install, easy to publish” can also mean “easy to exploit” when agents and tools sit inside real developer workflows. We unpack practical threat paths that show up fast in the real world, including: 📌 Indirect prompt injection through systems like tickets, issues, and other untrusted inputs 📌 Toxic flow patterns where untrusted content meets sensitive data plus outbound tools 📌 Malicious or poisoned MCP servers that nudge an agent into reading files or exfiltrating secrets 📌 Classic vulnerabilities in MCP server code (command injection, path traversal, SSRF, and more) 📌 Why “AI reviewing AI written code” can be helpful, but also inconsistent, and where deterministic scanning still matters If you’re building with MCP, approving MCP servers for internal use, or letting agents touch real systems, this conversation will sharpen your threat model.