Bedrock Web Search Needs Source Governance

On 4 August, AWS made Web Search for OpenAI GPT models generally available in Amazon Bedrock. The feature can incorporate current web information into answers and cite sources. For organisations, this is more than another model switch: web search creates a new, controllable data source in a production workflow.

What AWS provides

According to AWS, search runs server-side in Bedrock. It currently supports GPT-5.4, GPT-5.5, and GPT-5.6 Luna, Terra, and Sol through the Responses API. The model can formulate queries, receive titles, URLs, and snippets from Amazon’s index, and produce an answer with citations.

By default, the service uses the Bedrock web index and cache. However, AWS documentation draws an important distinction: direct access to the external web is governed by the external_web_access parameter and the bedrock-websearch:ExternalWebAccess IAM permission. Web Search is also currently regional only in us-east-1, us-east-2, and us-west-2. For DACH workloads, that is an architectural decision, not a footnote: the stated AWS service boundary does not replace an assessment of which data an application may send to which region.

From answer quality to source policy

Answer quality does not automatically improve just because a model can search. The key questions are which sources it may use, whether citations remain visible to business users, and when an insufficiently supported answer must be rejected. AWS says the model can identify a gap when results do not support an answer. That rule should be tested in the business process, not only in a demo prompt.

OWASP warns that agentic systems can suffer from excessive functionality, permissions, and autonomy. Applied to web grounding, this means not every application needs live web access; approved sources or an internal knowledge base may be sufficient. Where external access is needed, teams should treat it as a separate permission, rather than an implicit model capability.

What DACH organisations should define

Before a pilot, create a concise source policy: permitted data classes, allowed search and fetch modes, approved regions, citation visibility, answer retention, and business approval for critical decisions. Also test cases without reliable search results.

The useful question is not: can our model search the web? It is: which statement may it support with which source, and in which region?