decrypted · 22 november 2026 · vulnerabilities and patching · supply chain · ai and llm security
AI prompts were leaking through the cheapest route to the model
The AI routing service OpenRouter said on Sunday that it has removed three inference providers after researchers showed that customers' prompts were leaking. The evidence comes from Dragos, which spent five months feeding the AI supply chain booby-trapped prompts and watching who read them. The lesson for UK organisations is uncomfortable: the privacy setting you chose may bind nobody further down the chain.
How the trap worked
Think of marking banknotes with serial numbers, then watching where they turn up. Dragos wrote prompts containing unique web links and fake credentials, including fake Amazon Web Services keys, that no legitimate system would ever use. If anyone later opened the link or tried the key, someone had read the prompt. Dragos reports sending approximately two million prompts between July and November 2026.
What it found
Dragos found evidence that data from at least three AI providers was stolen. The earliest incident dates to mid-July, and none of the three knew until Dragos told them. In the largest, a couple of hundred fake keys were each sent in a single request, inside a few short windows in late October, and every request in those windows was stolen. Some keys came from requests rejected before any model ran, so the attacker was copying traffic at the provider's front door.
The thieves tested each key, sometimes within hours, then listed storage buckets, searched for stored secrets and tried to create their own users. The bait keys were used for tens of thousands of calls, and one was still in use three weeks later.
In two of the breaches, requests flagged ZDR were taken too. ZDR is zero data retention, the strictest setting, which tells a provider not to keep a prompt. Dragos put it plainly: a privacy flag does not help when the attacker is already inside.
No hacking required
On an inference marketplace, where sellers offer the same model at competing prices, Dragos found multiple resellers able to view and retain customer prompts using standard logging, without exploiting a vulnerability. Sellers only had to list a model at the lowest price, enable request logging and wait. In one case, someone on a home broadband line opened both secret links about a day and a half later and began guessing usernames. Dragos concluded that the leak follows the seller, not the marketplace as a whole.
What OpenRouter did, and the design lesson
OpenRouter says it off-boarded three providers, removed 125 endpoints from ZDR eligibility and notified all potentially affected customers. It says the findings showed no breach of OpenRouter itself, and that not all affected providers have posted publicly.
The Secure by Design reading, which is opinion, is about procurement. A price per million tokens buys you a route, not a promise. Prompts routinely carry client data, source code and credentials, so under UK GDPR you should be able to say who can read them. Choose your upstream provider yourself rather than the cheapest reseller, and plant your own fake keys to see who touches them. A fake key costs nothing.
Also this week
Fridge controllers, six unpatched flaws. Midnight Blue disclosed six unauthenticated buffer overflow flaws in the Danfoss AK-SM800A, a controller that runs refrigeration in supermarkets and cold stores. They affect firmware up to and including 5.3.2. CVE numbers are pending and disclosure to Danfoss is still under way. Midnight Blue says it found the bugs within an hour, that the application runs as root and that the device lacks stack canaries, a basic memory safeguard. It reports no attacks using these six, but says older Danfoss and Copeland flaws were exploited across Europe from 11 October. One of those older flaws was an undocumented daily access code that Claroty's Team82 reconstructed from the firmware, which Danfoss has since patched. A back door that anyone can recompute is the opposite of Secure by Design. Midnight Blue's advice is to remove internet exposure: use a VPN, a private APN (the access point name that connects a device to a mobile network) or a tight allowlist, and count controllers hiding behind integrators' 4G modems. UK cold-chain operators should ask their integrators that question this week.
A marketing tool, a customer list. Who Gives A Crap told customers in the UK, US and Australia that their details were accessed at a marketing software provider. Its notice, updated on 19 November, lists names, emails, phone numbers and mailing addresses, with no indication that payment details or passwords were involved. The company says it learned of the incident on 6 November, disconnected the provider and cannot name it. It says the provider used customer information to help analyse the effectiveness of its marketing, and TechRadar reports the company has no evidence of misuse so far but warns customers to be cautious. The design question is whether a tool that measures advert performance needed identifiable customer data at all.
Sources
- Dragos: Three breaches, endless leaks: Your AI supply chain needs canaries
- OpenRouter: Data Processing Incidents on Inference Providers
- Midnight Blue: Multiple RCE vulnerabilities in Danfoss AK-SM800A
- TechRadar: Danfoss refrigeration controller flaws found by Claroty Team82
- Who Gives A Crap: November 2026 Third Party Provider Data Breach notice
- TechRadar: Who Gives A Crap data breach
If you want help working out who can read your data, get in touch.
More like this
- Cursor's leaked supplier key ended in a booby-trapped Claude Code extension 15 november 2026
- N-able pulls its remote management platform offline after 30 alleged zero-days are published 10 november 2026
- A poisoned security update reached about 2,950 servers 7 november 2026
Get the next post by email: subscribe to Decrypted. Double opt-in, unsubscribe any time, or take the RSS feed.
Prefer to listen? Decrypted on Apple Podcasts, or paste the podcast feed into any app.