When NuLab began building AI-powered features for Backlog, its project management platform used by teams worldwide, the team quickly discovered that traditional security controls were not enough. Prompt injection attacks, unintended AI agent tool calls, and the risk of sensitive data exposure highlighted a new challenge: securing not just user input, but also the actions and outputs generated by LLMs. In this DASH session, Yuichi Watanabe, Principal Engineer at NuLab, shares how he built an open-source security middleware for the Vercel AI SDK that integrates with Datadog AI Guard. Learn how the middleware evaluates prompts and tool calls in real time, enforces centralized security policies before actions are executed, and provides end-to-end observability across AI workflows. Watch for practical insights on securing LLM applications, mitigating prompt injection risks, evaluating tool calls safely, and navigating production trade-offs such as monitoring versus blocking modes, fail-open versus fail-closed strategies, and SDK compatibility challenges.

 303      24