Web Search and Web Fetch, Now in Chat

May 13, 2026 · InfraScout Team

Assessments rarely stay inside the four walls of your tenant. Mid-review you reach for the outside world — a vendor advisory that explains a patch, the CVE behind a flagged version, the exact scope a Microsoft Graph permission grants, or what an obscure error code actually means. Until now that meant leaving the conversation, searching elsewhere, and pasting findings back. With this release, the AI can search the web and fetch specific pages without leaving the assessment, and every answer it draws from the web comes back with its sources cited.

This keeps the loop tight. When a finding needs context, the AI looks it up in place, links the supporting page next to the conclusion, and continues the review with that evidence on the record.

Research without leaving the assessment

Two capabilities work together inside the chat. The AI can run a web search to discover relevant pages, and it can fetch a specific URL to read it in full. In practice that means you can ask questions that previously sent you to a browser tab: "what does the latest vendor advisory say about this update," "is there a known CVE for this software version," "what does this Graph permission actually allow," or "what's the documented cause of this error."

The AI decides when to reach out based on the question. A query that can be answered from the assessment data stays local; a question that needs current, external information triggers a lookup. The result is woven into the same answer, alongside the rest of the review, so the context that led to a finding and the external evidence behind it live in one place.

Every claim cites its source

When the AI uses the web, the message carries a Sources footer listing the pages it drew from, each as a link you can open and verify. This matters for assessments specifically: a finding is only as trustworthy as the evidence behind it. Rather than asking you to take a web-derived statement on faith, InfraScout shows you exactly where it came from, so a reviewer can confirm the advisory, read the CVE entry, or check the documentation directly.

Cited sources are preserved with the message, so when you scroll back through a conversation or hand a session off for review, the links are still there next to the conclusions they support.

Admins control it per connection

Web research is governed at the level of each LLM connection, and it is available on Anthropic connections. A single master toggle on the connection decides whether the AI can reach the web at all. The capability ships enabled by default on eligible connections, so teams get cited research out of the box — but an administrator can switch it off for any connection where outbound web access is not wanted, and from that point the AI works only from your tenant data and the assessment at hand.

Because the control lives on the connection, you can run different policies side by side: a connection used for sensitive, air-gapped-style reviews with web research turned off, and another for general assessments with it on.

Domain allow and block lists

For teams that want web research but on a tighter leash, each connection also carries a domain policy. You can define an allow list, which restricts research to a named set of domains — useful when you only want the AI consulting, say, your vendors' advisory pages and Microsoft's documentation. Alternatively, you can define a block list to keep specific domains out while leaving the rest of the web open. A connection uses one approach or the other; you set an allow list or a block list, not both at once.

Each connection also carries a usage cap, so a single conversation can only make a bounded number of web lookups. This keeps research focused and predictable rather than open-ended, and it gives administrators a ceiling they can reason about per connection.

Domain entries are validated when you save the connection, so a malformed entry is caught immediately rather than failing silently later during an assessment. Combined with safe handling of redirects, the policy you configure is the policy that actually applies when the AI reaches out.

Try it

Open an assessment against an Anthropic connection that has web research enabled and ask something that needs the outside world — "is there a published CVE for this version," or "what does Microsoft's documentation say this permission grants." The AI runs the lookup, answers in context, and lists its sources in the message. If your connection is tuned to specific domains, the citations come only from where you allowed.

Questions or feedback? Reach us at info@infrascout.cloud.