Yes, but only if the assistant can read your scan data. Without a connection it can explain what a CVE is, but it cannot know which of your projects are affected. With a Model Context Protocol (MCP) connection, it can query your findings, and you can limit what it may read or change.
What MCP is
The MCP site describes it as an open source standard for connecting AI applications to external systems. An assistant such as Claude or ChatGPT can connect to data sources, tools and workflows through it. The site compares it to a USB-C port: one standard way to plug an assistant into many systems.
A server exposes capabilities. A client inside the assistant calls them. You decide which servers to add.
Without and with a connection
| No connection | Connected through MCP | |
|---|---|---|
| Knows your projects | No | Yes, within granted scope |
| Answers "which repos use this package" | Guesses | Looks it up |
| Can read current findings | No | Yes |
| Can take actions | No | Only if the server allows it and you permit it |
| Data leaves your tenant | Only what you paste | What the server returns |
| Main risk | Wrong answers presented as fact | Over broad access, injected instructions in data |
Safe setup steps
- Start with read only access. Reading findings is enough for most questions.
- Scope the connection to the projects the user can already see. The assistant should never have more access than its operator.
- Require approval for any tool that changes things, such as opening a pull request or merging.
- Treat returned data as untrusted text. A package description or advisory can contain instructions aimed at the model.
- Log tool calls so you can see what was asked and what was returned.
- Review the connection periodically and remove servers nobody uses.
Good questions to ask once connected
- Which of my projects contain this package and version?
- Which findings are reachable from entry points?
- Which findings have a fixed version available today?
- What changed in my risk since last week?
Always check an important answer against the finding itself. The assistant summarises; the record is the source of truth.
Try it yourself
Open the demo and see an assistant answer a vulnerability question using connected scan data. Notice which tool calls it makes.
Click through it at your own pace, or open it full screen.
Where Safeguard fits
Safeguard provides an MCP server for querying findings and an MCP Guardrail for controlling what connected agents may do. Start with a free scan to have data worth asking about.
Quick check
Will the assistant remember my findings afterwards?
That depends on the assistant's memory and data settings, not on MCP. Check the product's data controls before connecting sensitive data.
Is a read only connection enough for triage?
Usually yes. Reading findings, components and fix versions covers most triage. Keep write actions behind explicit approval.
Sources
- Model Context Protocol, What is MCP, modelcontextprotocol.io/introduction.
Last reviewed: 19 October 2026.
