Researchers used canary credentials in requests to AI routers like OpenRouter and Vercel, even under "ZDR" policies, and saw "privacy" often broken, keys stolen (Dan Goodin/Ars Technica)
Researchers evaluating intermediary services like OpenRouter and Vercel embedded canary credentials within outgoing requests to examine sensitive data handling under active platform use. The investigation revealed that user privacy protections were broken and keys were stolen during the routing process. These security compromises occurred despite systems operating under zero data retention policies. Consequently, the findings show that existing service safeguards fail to protect sensitive authentication artifacts.
Key points
Zero data retention policies do not guarantee that sensitive information remains safe on…: Zero data retention policies do not guarantee that sensitive information remains safe on artificial intelligence platforms.
Intermediary artificial intelligence routing services introduce exposure risks that can r…: Intermediary artificial intelligence routing services introduce exposure risks that can result in compromised credentials.
Embedding canary credentials provides a verifiable method for detecting unauthorized key…: Embedding canary credentials provides a verifiable method for detecting unauthorized key exposure across routing systems.
Sources checked by our analysis pipeline: 1
Evan Perez / CNN: |
Todd Bishop / GeekWire: |
Financial Times: |
Peter Hall / Scientific American: |
New York Times: |
Cristina Criddle / Financial Times: |
Financial Times: |
Dan Goodin / Ars Technica: |
The Information: |
George Steer / Financial Times: |
Bloomberg: |
@oracle: |
Cade Metz / New York Times: |
Cheng Ting-Fang / Nikkei Asia: |
Sophie McEvoy / GamesIndustry.biz: |
This is a Techmeme archive page. It shows how the site appeared at 3:10 PM ET, November 21, 2026.
The most current version of the site as always is available at our home page. To view an earlier snapshot click here and then modify the date indicated.
Industry observers and analysts are actively debating which groups will capture the true value created by advanced automated systems. Uncertainty persists over how the technology's broader impacts will distribute advantages across various sectors and end users. Resolving this issue remains critical as adoption deepens worldwide.
The open-source quantization compiler LookAtMyQuant has drastically reduced large language model memory requirements by combining trellis codes with proximal fine-tuning. This approach allows developers to run capable mixture-of-experts models locally on Apple silicon while maintaining benchmark performance that rivals or exceeds larger competing builds. Released under the MIT license, the compiler demonstrates significant potential for edge computing by maximizing quality within constrained RAM targets.