Devs, curious how you're handling sensitive data in your LLM applications.
retained text
**How seriously are you treating PII exposure in your LLM applications?** I’ve been thinking about something and wanted to get perspectives from people actually building LLM-powered applications. When we talk about sending sensitive data to LLMs, most conversations seem to focus on whether OpenAI, Anthropic, or another provider can be trusted with that data. But isn’t the bigger concern what happens to that data across the rest of the stack? A prompt containing someone’s name, SSN, medical history, or financial information could also end up in application logs, observability tools, error traces, conversation histories, agent memory, eval datasets, and downstream API calls. I’ve heard the analogy that developers wouldn’t knowingly put an API key into a prompt, even if they trusted the model provider. The difference is that you can rotate an API key. You can’t rotate someone’s SSN or medical history. I’m curious how developers here actually approach this. **Is PII entering your LLM pipeline something you actively worry about, or is it a relatively low-priority concern?** **Are you doing anything to prevent sensitive data from reaching the model or being retained across your infrastructure?** Or do you mostly rely on the LLM provider’s privacy and security controls? **If you’re already handling this, what’s your approach?** Redaction, tokenization, anonymization, self-hosting, contractual safeguards, or something else? **Where does this sit on your engineering priority list?** Is it something you’ve invested time and resources in, something you plan to address eventually, or simply not a meaningful issue for your use case? I’m particularly interested in hearing from people building production applications, especially those working with customer data. Also genuinely interested in hearing from developers who **don’t** see this as a significant problem. What’s your reasoning? Trying to understand how much of a concern this actually is in practice versus how much attention it gets in security discussions.
document fetched 2026-10-29 00:28 · canonical https://reddit.com/r/mlops/comments/1xhaw3h/devs_curious_how_youre_handling_sensitive_data_in