Field noteNov 22, 20266 min read1 source

The AI Finished its job. Three Days Later, Something Re-opened the Link.

In one minute

Ted Tschopp describes how a unique tracking link embedded in a test prompt can reveal an unexpected later request—and why that signal does not, by itself, prove a vendor breach. For teams using AI services, the practical lesson is to connect privacy promises to incident records, supplier explanations and a clear decision about whether sensitive work should continue.

What changed: The article reports Dragos research finding that planted tracking links in AI requests were reopened later, raising questions about providers’ privacy controls and data access.

What happened

In an essay published November 21, 2026, enterprise architect Ted Tschopp asks readers to imagine a company testing an AI service before sending it real factory repair reports. The security team places a unique link in a fabricated report and sends it for summarization. The linked page records when someone or something opens it. The AI returns a summary, and the test appears complete—until the link is opened three days later.

Tschopp uses this hypothetical event to frame an investigation, not to claim that a particular company experienced it. A later request shows that the link was accessed, but it does not establish who accessed it, how they obtained it, whether they were authorized, or what information was retained. A system might have kept only the URL or queued it for later processing; the event does not prove that the full report was stored.

The essay situates the scenario alongside two public accounts. Tschopp says Dragos published research on November 19 involving approximately two million AI requests containing unique tracking information. Dragos reported compromised providers, access by sellers of AI services, and behavior that appeared inconsistent with privacy settings. Some planted information was used in attempts to sign in to controlled test accounts. The essay also notes a limitation: detecting use of planted information did not automatically reveal where it escaped, and some companies and systems remained unidentified. Dragos’s findings do not establish what percentage of enterprise AI requests are exposed or show that every supplier has the same problem.

Tschopp describes the route a request can take: an application may send a prompt through a model-routing service to a model provider. Each intermediary raises questions about what it can see, do and retain. The essay distinguishes three promises that are often conflated: not training on customer information, not retaining it after a request, and restricting who may read it. It points to documentation from OpenAI and Anthropic on data controls and retention, and to OpenRouter’s documentation and data-processing agreement, which distinguish its request-retention controls from every additional enabled tool and distinguish provider-policy selection from a guarantee of provider compliance.

For an unexpected link visit, the suggested investigation includes preserving timestamps and request records, checking whether the link was copied elsewhere, identifying the system that made the request, and comparing the explanation with the service and configuration the company actually purchased. The supplier should receive a specific request containing the request identifier, event time and expected behavior. Meanwhile, the company should decide whether to pause sensitive use. Tschopp cites Phala’s November 10 preliminary incident report as an example of separating confirmed misuse in sampled requests from a wider period customers were advised to treat as potentially exposed. The essay says that report does not establish whether Phala’s incident was among Dragos’s unnamed cases.

What the archive adds

New Runtime’s earlier coverage connected AI systems to consequential operational risks: on November 17, 2026, Federal Reserve Governor Michael Barr warned that agents could accelerate bank runs. Tschopp’s essay is a new event and a distinct, practical treatment of supplier-data investigation rather than a restatement of that financial-stability warning.

Trend signal

This supports agent security moving to runtime boundaries: a privacy promise or prompt setting cannot establish what happened across an application, router and model host. Teams need evidence about actual access and retention, plus an operational owner who can decide what to do while the evidence is incomplete. The essay is about AI services broadly, not only autonomous agents, so the connection is relevant but not a direct demonstration of agent runtime controls.

Why now

The essay points to Dragos’s November 19, 2026 research and Phala’s November 10 preliminary incident report as timely examples of tracking signals and response under uncertainty. Those accounts make the hypothetical investigation more concrete without establishing that the same incident or cause links them.

Connections to other events

New Runtime’s November 17, 2026 coverage of Michael Barr’s warning about AI agents and bank runs also concerns risks that can unfold faster than institutions can respond. Tschopp focuses on a different operational scale: tracing information through suppliers and deciding whether to suspend sensitive use while an explanation is pending.

Why it matters

Teams often assess AI privacy through contract language, configuration screens and vendor assurances. Tschopp’s scenario shows why those are not substitutes for an incident process. A link opening can establish that a request occurred, while leaving identity, authorization and scope unresolved. Treating it as proof of a breach risks a false accusation; dismissing it because the service promised not to retain data risks continuing exposure.

The useful practice is to make the response part of deployment: preserve request and event identifiers, know which suppliers handled the prompt, assign someone authority to coordinate security, IT, procurement and privacy, and set a deadline for a supplier explanation. Decide in advance what kinds of data justify pausing work. Keep investigation records limited and access-controlled; retaining every conversation indefinitely creates another sensitive store. This turns privacy commitments into questions a team can investigate and act on, rather than labels it can only trust.

Limits and uncertainty

The tracking-link scenario is hypothetical, and the essay does not identify a company whose AI prompt was reopened after three days. A link visit cannot show who made the request, what data was retained, or whether access violated a contract. Dragos’s reported sample does not establish an exposure rate for enterprise AI use, and the essay says some implicated systems were not identified. It also does not establish that Phala’s reported incident was connected to Dragos’s findings. A canary that never triggers cannot prove that information was not copied or retained by another means.

Skill to build now

Run a controlled test with fabricated information and a unique tracking URL before approving a service for sensitive work. Record the request identifier and send time, then define who receives an alert if the URL is opened later. Practice the response: check whether the URL was copied elsewhere, send the supplier the identifier and event time, and name who can pause sensitive use pending an explanation. Do not put real confidential data into the test.

Reader takeaway

A tracking link is an instrument for detecting an event, not a verdict about a supplier. The value comes from pairing that signal with request records, a specific supplier question and a named decision-maker who can limit exposure while the facts are checked.

Evidence / sources

  1. 1tedt.org/The-AI-Finished-its-Job-Three-Days-Later-Something-Re-opened-the-Link
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicAnthropicExplore the anthropic topic hub.
  2. 02archiveField NotesOpen the latest editorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract