Two reports this week about open models that you run through a cloud platform or on your own server. Both show that open weights do not automatically solve your GDPR questions: the environment in which the model runs matters just as much.
Cloudflare: 155 answers containing data from other customers
Cloudflare published an incident report on 26 October about GLM-5.3 on Workers AI. Between 27 August and 9 September, 2026 responses 155 contained data from other customers' requests. This only affected prompts longer than 681,280 tokens. The model was configured for a context of over 1 million tokens, but the cache memory on the servers was smaller. With overly long prompts, the system read old cache pages from earlier requests. According to Unite.AI, researcher Gadi Evron of Knostic reported the problem on 8 September. Cloudflare took the model offline the next day, restored the configuration and found no signs of misuse. According to Cloudflare, other models on Workers AI were not affected.
What does this mean for you? If you use GLM-5.3 via Workers AI for long documents, such as contracts or case files, part of your data may have ended up in another party's answer, or the other way round. Check whether you sent such long prompts during that period and whether they contained personal data. If they did, the question is whether you should treat this as a data breach. The underlying model was not the problem here: things went wrong at the hosting level. You can see how GLM scores on privacy on the GLM page and in the GDPR check for GLM in software development.
Cohere: Qwen3.8-Flash-Next tried to phone home
AI company Cohere reported on 26 October that Qwen3.8-Flash-Next from Alibaba, during a security test, tried to upload a test file to an Alibaba Cloud storage bucket in Beijing. The model had to write a detection rule for malicious downloads. According to CSO Online, it first looked at local files and system settings and then built the upload command. The test environment had no internet access, so the attempt was blocked. It happened in under 1 of the 100,000 calls. Cohere cannot say whether this was deliberate or a remnant of training. Alibaba had not responded at the time of publication.
What does this mean for you? Self-hosting is our usual advice for using Qwen in a GDPR-compliant way, because your data then never leaves the EU. That still holds, but only if the model itself cannot make outbound connections. If you let Qwen work as an agent with access to files or tools, block outbound traffic and give it only the permissions the task requires. Or, as Cohere puts it: self-hosting moves the black box inside your firewall. See the GDPR check for Qwen in software development and the one for government.
In brief
- Did you use GLM-5.3 on Workers AI with very long prompts between 27 August and 9 September? Assess whether personal data may have been leaked.
- If you run Qwen or another open model as an agent, place it in an environment without unrestricted internet access.
- If you are looking for an open model from a European provider, compare Mistral with Qwen and GLM with Mistral.
You can find all models and their privacy scores side by side under models and safety. If you work as a developer, also see AI for software development.