Security Docs That Follow the Code: Dropbox’s MCP+Dash Combo

Why Security Docs Get Lost

You plot a course, check the charts, set the sails—but the wind changes. In software, security requirements are decided during design, but by the time code is ready for review, that context is a distant memory. Dropbox wanted to fix that. Their approach borrows from sailing: keep the navigator’s notes right next to the helm.

They hooked up their internal knowledge base, Dash, to the Model Context Protocol (MCP). Now threat models and design docs pop up directly in the code review screen. No more digging through old Confluence pages. No more relying on someone’s memory of a meeting six weeks ago.

When Documents Drift Away

In a lot of places, threat models and security requirements live in separate silos. They’re written once, then forgotten. As code changes, those docs get stale. Reviewers are left trying to figure out the “why” behind a change by chasing down old emails. It’s like navigating with a star chart that’s slowly spinning out of alignment.

Dropbox’s system uses Dash as a centralized index, respecting access controls while making knowledge searchable. MCP gives AI systems a standard way to pull that context into developer tools. The result? When you open a pull request, the relevant security notes are right there, next to the diff.

Not Another Bot That Says “Looks Good”

This isn’t about automating security reviews. It’s about making sure the context is visible. “We’re not replacing security reviewers,” says Ishan Mishra, the engineering lead. “We’re making sure agreed-upon requirements are harder to miss.” The system analyzes the code changes, then uses MCP to fetch related threat models from Dash. They appear alongside the diff, so the reviewer sees the design intent without leaving the code view.

Why MCP Instead of a Custom Plugin?

Building one-off integrations for each workflow would be a maintenance nightmare. MCP gives a standardized way to plug Dash in as a context provider. “We can iterate quickly,” Mishra notes. The review agent doesn’t need to know where information lives or how retrieval works—it just asks for context, and Dash handles that.

The architecture is reusable beyond security. The same pattern could apply to privacy, compliance, API governance, or design review. It’s a pattern that connects any organizational knowledge to the point of implementation.

Keeping Trust: No False Alarms

One risk with AI-assisted review is that devs might assume the AI has “checked” the code. Dropbox is careful. “We don’t treat this as a source of truth,” says Mishra. Every finding has to be traceable. You can click back to the specific security requirement, its source, and the corresponding code lines. If the system can’t provide evidence from both the requirement and the implementation, it stays quiet.

That restraint is key. Developers already get bombarded with automated feedback. False positives erode trust fast. So the system errs on the side of silence unless the connection is clear. It’s designed to assist, not to certify.

Reliability Means Relevance, Not Just Speed

Scaling this wasn’t about making it fast. The hard part is finding the right context. “In a large org, design docs and code aren’t always directly linked,” Mishra says. “Keyword search isn’t enough.” They use semantic retrieval, but only if the results are genuinely relevant. If a finding isn’t tied to the change, it’s noise.

They evaluate continuously, feeding back into the retrieval and reasoning models. Reliability here means the output is specific, actionable, and grounded in the code. Not just “hey, this looks like a security thing.”

Speed vs. Depth: Finding the Sweet Spot

During a live code review, you don’t want a research paper. You want a few key facts at the right moment. The system retrieves just enough to understand design intent, then presents concise, evidence-based notes. “We avoid showing every weak signal,” Mishra explains. If it can’t clearly connect a requirement to the current code, it says nothing.

Latency matters because code review is interactive. Developers are waiting. But they’d rather wait a second longer than get a bogus warning. The balance is tricky, but they’ve made it work by focusing on clear, defensible connections.

What Other Teams Can Learn

The biggest takeaway from this project: AI agents become far more valuable when they build on the organization’s past decisions, not just the current task. “Many AI coding workflows focus on generating or reviewing code in isolation,” Mishra says. “That’s useful, but it misses the bigger question: why is this code being written?”

This pattern goes beyond security. Privacy requirements, API conventions, architectural decisions—AI can connect implementation to accumulated knowledge. The goal isn’t just to make developers faster. It’s to help organizations preserve and apply the knowledge they’ve already built, at the moments when it matters most.

For sailors, a good navigator doesn’t just tell you where you are. They remind you where you’re headed and what you agreed to avoid. Dropbox’s MCP + Dash integration is like that: a quiet voice that keeps the journey on course, even in rough waters.