If your business routes chatbot, coding assistant, or customer support traffic through an AI gateway and trusts the "Zero Data Retention" (ZDR) label to keep prompts private, new research should change how you think about that setting. It did not work as promised.
What happened
OpenRouter, a popular gateway that routes AI requests to many underlying model providers, has off-boarded three inference providers from its platform, stripped 125 endpoints of their Zero Data Retention eligibility, and notified customers who may have been affected.
The trigger was independent research from industrial cybersecurity firm Dragos. Between July and November 2026, Dragos seeded roughly two million AI prompts with canary credentials: real AWS access keys tied to empty but fully logged accounts, fake passwords, and unique tracking URLs. The goal was simple — if someone other than the intended AI model ever read those prompts, Dragos would know, because the bait credentials would get used or the tracking links would get clicked.
They did. According to reporting from The Hacker News, attackers had been quietly reading customer prompts inside three providers for months — and none of the three providers knew until Dragos told them. Some of the compromised requests were explicitly flagged as Zero Data Retention, the setting meant to guarantee a prompt is never stored or read after processing.
How bad was it
In the largest case, attackers took around 200 of the bait AWS keys and used them for tens of thousands of AWS API calls, according to analysis from mlq.ai — sometimes testing a stolen key within hours of it appearing in a prompt. One unnamed Chinese AI lab fetched the secret tracking links from more than half of all seeded prompts sent to its models, and about two-thirds of the ones marked "no data collection" or Zero Data Retention, per the same Hacker News report.
In plain terms: whatever anyone pasted into a prompt — passwords, API keys, customer records, source code, contract drafts — was potentially exposed to a third party who had no business reading it, regardless of the privacy setting selected.
Why this matters for your business
Many Swiss SMEs now use AI tools daily: coding assistants, customer support bots, document summarizers, internal chat tools. Few check which underlying provider actually processes the data, or whether a "privacy" toggle is enforced end-to-end rather than just promised in a dashboard. This incident shows that a compliance checkbox is not the same as a technical guarantee.
What to do today
- Inventory your AI tools. List every chatbot, coding assistant, and automation that sends data to an external AI model, including ones employees adopted on their own.
- Check your gateway and provider. If you use OpenRouter or any multi-provider AI gateway, review the notification from OpenRouter to see if you were flagged as potentially affected.
- Rotate anything ever pasted into a prompt. If an employee ever pasted an API key, password, or credential into any AI tool, treat it as compromised and rotate it now.
- Stop pasting secrets into prompts. Make it a written policy: no passwords, no API keys, no customer personal data goes into an AI chat window, regardless of what privacy setting is enabled.
- Prefer providers with verifiable, audited data handling over ones that simply offer a toggle, and ask vendors directly how ZDR is technically enforced, not just labelled.
The bigger picture
This is the latest reminder that AI adoption is outpacing the security controls wrapped around it. We are tracking these incidents daily because the pattern repeats: a convenient feature, an unverified trust assumption, and months of silent exposure before anyone notices. For Swiss SMEs without a dedicated security team, that gap is exactly where damage happens unnoticed.
If you want a second pair of eyes on which AI tools your business uses and how data actually flows through them, The Mejlerö Company (Flitz.ai) in Engelberg offers a free network analysis and a free one-hour consultation for businesses in Obwalden, Nidwalden, Lucerne and Zurich. Beyond support calls and endpoint security basics, we also help SMEs automate parts of their business securely — worth a conversation if AI tools are becoming part of your daily workflow and you want that done safely rather than by accident.