A prospecting agent should not open every link it finds

OpenAI shows how an agent can leak private context through a URL without ever putting that information in its written answer.

OpenAI published a description of its URL-safety controls for agents in January 2026. The controls address a specific risk: an attacker can hide private information inside a URL and induce an agent to fetch it.

The risk can sit inside a link preview, image load or other background request. The agent does not need to write the private information into its answer for the destination to receive it.

OpenAI explains that websites log requested URLs, so a URL containing an email address can expose it to the destination. It identifies prompt injection in web content as one way an attacker can trigger the request.

Its safeguard tests whether an independent crawler has already observed the exact URL on the public web. A match can be fetched automatically, so an unmatched URL requires a warning or explicit user action.

The control prevents URL-based data exfiltration, but it does not prove a page is truthful or safe in every other sense. It sits alongside monitoring, prompt-injection defences and red-teaming.

This is a product-security explanation from the team operating the control, not an independent study of its effectiveness. The mechanism is documented, but the source does not publish a measured reduction in attacks.

Prospecting agents work across mixed-trust material. A company page, press release, LinkedIn profile and prospect-shared document should not receive the same default permission.

The important design choice is the boundary between reading information and taking an external action. An agent can collect public facts at volume, while a person approves any link or tool action carrying account context outside.

That boundary matters before sending an email. An enrichment step that leaks a prospect's details is a serious research failure, because the recipient cannot correct it.

The boundary between reading information and taking an external action belongs in the workflow, not in a hopeful prompt.

List every external action your research workflow can take, including opening links, downloading files, loading images and calling enrichment tools. Then decide which inputs may trigger each action without a person.

Treat links from public search results differently from links assembled from CRM fields, emails or shared documents. The second group carries more context and deserves a review step before an agent follows it.

Keep the agent's permissions narrow while you test it. Read-only research is easier to contain than an agent that writes to a CRM, contacts a third party or changes a record.

Mapping which of those external actions already run without a person is where the Diagnostic starts.

Sources

How we choose and label these: our sourcing standard.

  • Keeping your data safe when an AI agent clicks a link →

    OpenAI · Primary source · link checked 4 Aug 2026

    OpenAI's account of one URL-safety control in its agentic products. It describes the mechanism and its intended limit, but does not provide an independent effectiveness study or an attack-reduction statistic.