Skip to content
Interface language: English
Audit / Signal

A link in an answer can also send data outside the company

Trace requests caused by images, links and previews in AI answers, then design a test with synthetic data.

Article language: English

A person holds a red comb against a long dark braid beside a bright window.
In this article

A small image appeared in an assistant's answer. While an employee read a summary, the browser fetched the illustration from an external server. The path or query parameters sent to the recipient contained information from the material being analysed. Nobody attached a file to a message. A recipient appeared simply because the result was displayed.

This is a hypothetical sequence in a translation office. The assistant organises a project schedule using approved documents. It must present the answer in an internal panel without passing content to new services. The team has restricted the agent's sending tools, but has not yet checked what the panel does with a generated image or address.

The issue covers the whole path of the answer. A model produces text, but another component may turn it into a link, fetch an image or prepare a page preview. The assessment must therefore include where the answer is displayed and its connections. “The agent sends nothing” is too narrow if a browser or preview server also operates beside it.

First identify who makes the request

In our office the answer goes to a conversation panel. It can show plain text or interpret formatting. A renderer is the component that turns the markup into the visible result. If it recognises an image specified by an external address, determine what happens next in the actual implementation. Panels do not all support the same elements.

Microsoft describes the risk of uncontrolled handling of AI output, including an image request triggered when a result is rendered. It recommends treating the answer as untrusted before passing it to subsequent components.

For one link, map who recognises the address, who initiates the fetch and where the response appears. It might be the employee's browser, a thumbnail service or an agent tool visiting a page. Each path has its own configuration and owner. Do not treat them as a single general “links” function.

A preview server of your own does not end the question of the recipient. If the intermediary sends a request that reaches the external service, the recipient may receive data in the transmitted path or query parameters. Check attempted requests, blocks and actual receipt in the traces. The component making the connection may change, but data placed in the requested address still needs control. Hiding the employee's computer address is not the same as removing content from the destination URL.

Record when the action starts. An ordinary link deliberately opened by a user is a different sequence from an automatic preview or an image fetched when the answer is displayed. In some products, expansion of the result or hovering over the link may be the relevant moment. Acceptance should cover functions actually available in the panel, without inferring them from appearance alone.

Connect the assessment of the agent sandbox and its network limits (in Polish) to this map. If the agent has no internet access but the panel opens addresses in its own layer, a passing agent block does not describe the panel's behaviour. State which component each control covers.

The address is part of the data sent

In the translation office, the project name and deadline belong to the task material. If they appear in the destination address of an external fetch, they do not cease to be information because they were written into the URL path or query parameters. The fragment after # is not sent in an ordinary request, though an application may separately read and transmit it another way. During acceptance, inspect the whole address, not just the server name and the label visible to the employee.

A short label such as “source” can lead to a much longer destination. A thumbnail can look like an ordinary logo. Neither view automatically shows which parameters were attached. If opening the destination requires a user's decision, the interface needs a way to identify it before the action. Approving the appearance of the answer is not informed approval of every background request.

A consent question must precede the action it concerns. If the panel fetches a thumbnail first and only then asks “Open the external page?”, the confirmation concerns a different, later step. It cannot undo the earlier fetch. If that is the agreed scope of work in the translation office, the decision to contact a new destination must precede that specific action. Add refusal to the test and inspect the trace from the answer's first display, not only from a button press.

Do not confuse this control with the header carrying information about the referring page. MDN describes Referrer-Policy as controlling information sent in the Referer header. It does not describe removal of data from the destination URL.

The practical implication for our process is straightforward: limiting referring-page information does not replace assessing the address the application itself prepared to fetch. These should be separate items in the evidence. Otherwise the team may correctly check one control and mistakenly conclude that it has accounted for all traffic generated by the answer.

Also check redirects if the mechanism in use follows them. A known starting address is not automatic proof of the identity of every later recipient. The implementer should state how the panel or intermediary handles a destination change and which restriction remains active. Reading the first server's name from the visible link is insufficient.

A signature on an event triggering a process (in Polish) also does not authorise fetching every address in its content. A genuine message from a company service can contain text added by an external person. The origin of delivery, the origin of a link and permission to transfer data have different bases.

Let people read the result without automatic contact with a new destination

The process owner should first determine whether an external image or preview is needed at all. In a translation schedule summary, the useful information is deadlines and dependencies. A decorative illustration may add nothing to the task. Forgoing an automatic fetch can then be simpler than maintaining an elaborate exception for decoration.

Aeneia conceptual photography

A link to an approved source may still be needed to check the content. The design should distinguish displaying an address from visiting it and offer an agreed way to open it. The user should be able to recognise what will happen and which component will do it.

OWASP describes Improper Output Handling as insufficient control of output before subsequent use. It points to validation and encoding appropriate to the context. One generic “cleaning” operation does not describe every way the answer will be used.

For the panel, agree with the implementer which formatting elements and external resources are allowed. An image, ordinary address and fetched preview may follow different rules. Their description should concern the version in use. Disabling active HTML does not by itself answer the question of the other supported elements.

In a browser one layer may be CSP, Content Security Policy: rules specifying the resources a page may load. The img-src directive specifies allowed image sources. Its name and scope do not automatically block every navigation or requests made by a separate preview server.

Acceptance therefore needs to identify where the policy actually applies. If an external image was blocked in the browser, that result concerns that request and configuration. An intermediary fetching a thumbnail has its own path. Do not extend one correct result to every component merely because the user sees them in one panel.

A test with a harmless marker

Examining this path does not require confidential documents. With the implementer, prepare a closed environment, synthetic material and a controlled test receiver. The marker should be explicitly fictional, unlike customer data and easy to find in the test record. This is a proposed control, not a claim that the article performed a test.

First define the normal task result: an answer available in the panel without transferring the marker outside the specified scope without agreement. Then exercise the agreed display and opening methods. For each, record whether the address only appeared in text, whether a connection was attempted, whether the request was blocked and what the controlled receiver obtained.

Visible address text does not itself prove transmission. Conversely, no image on screen does not prove that no request occurred. The fetch may have failed or the image may remain invisible. Assessment should use the appropriate network trace, not only a conversation screenshot.

Add ordinary, permitted control material. If the panel blocks every reference to a source, it may prevent necessary verification. In our office the goal is to retain a link to the correct project instruction without automatically disclosing content to another recipient. Test the effectiveness of the restriction and the usability of the approved path separately.

Keep the material version, answer before display, panel configuration and actual steps in the evidence package. The implementer should indicate where the request or block was recorded. The agent tool's own log may not include a connection initiated later by the browser. This need not be a fault in that log; it may mean the wrong place was inspected.

Compare the observation with the answer before display. This helps identify where the address came from; if several layers intervene, their traces are also needed. If receipt was confirmed, use the available traces to identify which component passed the marker. For a blocked attempt, record the absence of confirmed receipt and the available traces without assuming delivery. A marker in the final log does not explain the whole cause. The note should connect the text version to panel behaviour and the recorded request. It should not attribute to the model an action performed by another part of the application.

An answer can change its environment after being saved

A result read without fetches in one panel may be copied into a message or document. On the way, the way its content is interpreted changes. An AI-generated HTML or SVG file (in Polish) requires control of active content and opening conditions. Here we start earlier, with an ordinary conversation answer that may later become such material.

An Office document with links and external resources (in Polish) needs its own separate acceptance. There is no need to assume every export is dangerous. Identify the format and use path actually assessed. Saving a file does not automatically carry the panel's restrictions into the recipient's application.

After changing the panel, repeat the case that revealed the unauthorised request. Also check that the correct instruction can still be opened through the permitted path. If the renderer, thumbnail service or export method has changed, the earlier evidence may not cover that version. Acceptance should identify current components, not just the assistant's name.

Remember previously saved conversations as well. An employee may open an old answer in the new panel. Agree with the implementer whether the changed rule covers this return and exercise it with test material. Accepting only new generation is insufficient. For the office, an old schedule is still a working document, and displaying it again should follow consciously established rules.

Prompt injection in documents (in Polish) explains where a foreign instruction can originate. This review determines who later makes the network request. Separating the questions allows a specific effect to be restricted even when flawless interpretation of every future model answer cannot be promised.

At Aeneia Audit, discuss the scope of an operational review of AI use and data flows, and whether a separate specialist security assessment is needed. Technical protection of the renderer and network requires the appropriate implementer and separate confirmation. For the translation office, the goal is a summary that can be read and checked against the stated source. An external image should not quietly add a new recipient of the schedule.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear