Every AI agent processing meeting transcripts, support tickets, or emails is funneling raw personal data through cloud APIs. Names, salaries, medical details, social security numbers - all of it. VeilPhantom is a new open-source Python SDK built to fix that specific problem.
The approach is straightforward: VeilPhantom scans text on your device, detects personally identifiable information (PII), and replaces it with generic tokens like [PERSON_1] or [AMOUNT_1] before anything leaves your machine. The LLM processes the sanitized text, returns a response with those same tokens, and VeilPhantom swaps the real values back in locally. Your cloud provider never sees the sensitive data.
The more interesting result is what happened to output quality. Across 98 benchmark scenarios spanning 8 industry verticals using Claude Haiku, accuracy actually went up with tokenized inputs. That runs counter to what you might expect - stripping context usually degrades performance. But it turns out LLMs can reason just fine about [PERSON_1]'s salary review without knowing the person's actual name, and removing the noise of real PII seems to help the model focus on the task.
For anyone building AI workflows that touch customer data, employee records, or medical information, this fills a real gap. Most teams either send everything to the API and hope their provider's data policies hold up, or they build custom regex-based scrubbing that misses half the edge cases. An on-device detection layer that handles tokenization and rehydration as a single SDK call is a more practical middle ground.
VeilPhantom is available now as a Python package with the SDK documentation on its site.