AI Supply Chain Security Starts Before the Model Answers
Dragos found decoy credentials in use after AI requests were rejected. Trace where your prompts go, what zero retention covers, and what a canary alert tells you.
Ernest Bursa
The provider rejected the prompt before the model ran. Someone still used the decoy credential inside it. Dragos reported that some of the test credentials used in its largest observed incident had appeared only in rejected requests. If your team treats a failed request as one that never left the building, that distinction matters. Dragos’s research
Think about a vulnerability report arriving on a Friday afternoon. An AI assistant could pull out the affected component and suggest which team should take a look. The original report might also contain a private attachment, a reporter’s contact details, or a credential copied into a log. By the time the assistant returns a category or an error, whatever you put in the prompt has already been sent. Your decision about where that data can go belongs before the request.
The model name leaves out the route
Dragos sent roughly two million prompts containing unique test markers to widely used AI services between July and November 2026. Its researchers reported evidence of exposure at three hosting providers. They also observed marketplace sellers with access to prompt logs and some links being fetched in circumstances inconsistent with expected handling. These were different findings in the services Dragos tested. Each calls for a different investigation. Dragos’s research
When you use an AI gateway, the model you request and the provider that serves it can be separate choices. OpenRouter, for example, documents separate controls for model selection, provider selection and fallback. Your triage request might reach a different endpoint when the first provider is rate-limited. That can be a sound availability decision for public text. For a private report, the fallback needs to be approved to see the same fields as the primary endpoint. OpenRouter’s routing guide
Zero Data Retention (ZDR) concerns what a provider keeps after processing your prompt. Under OpenRouter’s ZDR-only routing, endpoint eligibility depends on the downstream provider’s stated policy. The agreement distinguishes eligibility from a guarantee of compliance. Its guidance also treats external tools such as web search separately from inference retention. Ask where the prompt goes while it is being processed, what persists afterward, and whether the answer changes when a tool is enabled. OpenRouter’s data processing agreement, §2.4 · OpenRouter’s retention guide
A canary alert gives you an event to explain
A canary token is a marker designed to alert you when someone or something interacts with it. A web canary is a unique URL that alerts you when it is requested. An authorized link scanner or preview can fetch it. The alert gives you a time and a marker to investigate. Finding the system that fetched it, the path it took, or how long it persisted takes other evidence. A quiet canary tells you only that you observed no interaction that triggered an alert during the test window. Thinkst’s web canary documentation
Start with a synthetic vulnerability report in a workflow you are authorized to test. Use a test URL that contains no sensitive information, and keep the report fictional. Confirm the alert reaches your team with a known test fetch, and record which approved components you expect to fetch links in normal operation. Save the test’s timestamp, request ID, route configuration and retention setting so the alert can be connected to a particular request. Repeat the check when the route or enabled tools change.
If a marker surfaces unexpectedly, whoever is on call should be able to compare the alert with request metadata from the gateway and tools. If they still cannot explain it, give the provider your request ID and timing through its security channel. Keep the real report in your restricted case system while you investigate. Decide in advance who can pause the workflow so the next report goes to a human. NIST’s incident response guidance
Decide what leaves before Friday
If the assistant is choosing a team, start with the fields it needs to make that choice. A component label and a short, reviewed description may be enough. Keep the attachment and reporter details with the case. Document the approved gateway, endpoints and tools, including the fallback, beside the list of fields they may receive. This gives you something to review every time the integration changes. NIST’s supplier-risk guidance for generative AI
When the next Friday report lands, the on-call engineer should already know what the assistant can see, where those fields may go, and where the report waits if the approved route is unavailable. You can rehearse that last handoff in the same vulnerability reporting drill you use to check that a report reaches a human.
Related articles
Try Kit for 30 days.
Hiring, security reports, and training in one account, for teams where none of it is a full-time job. Free for 30 days, card required. Cancel before it ends and you pay nothing.
Get started free