Skip to main content
← Learning Center
Data Privacy

AI Training Data Risks

Every prompt you send to an LLM may be used to train future models. Your proprietary data, PII, trade secrets, and source code can become part of a training corpus accessible to millions of users. See how Transmute prevents this, technically, not just with an opt-out checkbox.

Training Data Flow Simulator

Toggle Transmute ON/OFF to see what reaches the training pipeline

Prompt Sent to LLM Provider
Please analyze this patient record and suggest treatment options: Patient: Sarah J., DOB 05/12/1982, SSN 421-88-9012 Insurance: BlueCross ID BC7890123456 Email: sarah.jones@example.com Current medications: - Lipitor 20mg daily - Metformin 500mg BID Lab results from 06/20/2026: HbA1c: 7.8%, LDL: 142 mg/dL, eGFR: 52 Recent complaint: persistent headache behind eyes, occasional blurred vision. BP 148/92. Additional context: patient's employer (Acme Corp, account #ACM-88421) requires quarterly health summaries. Patient's auth token for portal is acme_prod_x8k2.
Data Flow Path
Your Machine
Data originates here
→
Transmute Proxy
Passthrough →
→
LLM Provider
Receives prompt
→
Training Data
CONTAINS PII!
Without Transmute, the LLM provider received a prompt containing raw PII including a patient name, SSN, insurance ID, employer name, and an active authentication token. This data will be retained per the provider's policy and may be used to train future models.
Policy Explorer

Provider Training Policies

Each LLM provider has different training data policies, and they can change. Click any provider to see the detailed policy, the risk, and how Transmute provides defense-in-depth regardless of the provider's current terms.

Real-World Scenarios

How Training Data Leaks Happen

Three real-world patterns where sensitive data enters LLM training pipelines. Expand each to see exactly how the leak occurs, and how Transmute prevents it at the technical level.

Stop Training on Your Data, Permanently

Transmute runs locally on your machine, redacting sensitive data before it reaches any LLM provider. Your prompts still get answered, but your PII, matched secrets and sensitive values are replaced before covered requests are forwarded.

Explore Transmute Talk to Us →
FAQ

Common Questions

It depends on the provider and your plan. OpenAI's ChatGPT Free and Plus tiers train on user conversations by default. This includes any data you paste into the chat. OpenAI's API and Enterprise plans do NOT train on data by default (since March 2023). Claude consumer model-improvement settings differ from commercial API terms; review the applicable settings and contract. Google's Gemini web app may train on consumer data; Google Cloud API does not. The key reality: policies change, and once data leaves your machine, you're relying on a provider's promise, not a technical guarantee.
Research has demonstrated that LLMs can memorize and regenerate training data verbatim. A 2021 study by Carlini et al. showed that GPT-2 could be prompted to output exact training sequences including names, phone numbers, and addresses. A 2023 study extended this to larger models. The risk is not theoretical, the bigger the model and the more unique the data, the higher the memorization probability. Transmute eliminates this risk entirely by ensuring the original sensitive data never enters the training pipeline.
You can, and you should. But opt-out checkboxes are policy controls, not technical controls. They protect you from intentional use, not from data retention for abuse monitoring (common across providers), accidental inclusion in training runs, or policy changes. A provider can update their data usage terms with 30 days notice, and your historical data is already in their systems. Transmute provides a technical control: the data is redacted before it leaves your machine, reducing exposure of matched values while provider training and retention policies remain relevant.
General data leakage is point-in-time, someone intercepts or views a specific prompt. Training data risk is persistent and multiplicative, your data becomes part of a model that can be queried by millions of users indefinitely. A leaked prompt exposes data once. A trained-on prompt exposes data every time an attacker successfully extracts it from the model, which can happen months or years later. Training data risk compounds over time as more prompts accumulate in training sets.
API keys and authentication tokens are the highest-value target, they're instantly actionable. Customer PII is second. It enables identity theft, spear-phishing, and account takeover. Source code and proprietary algorithms are third. They can be used by competitors or sold. Internal strategy documents and financial projections round out the top tier. Transmute's filter packs are explicitly designed to catch all four categories at the request level.