How does sensitive data leak into LLMs?
Through prompts people type or paste, files they upload, documents retrieved to ground answers, and training or fine-tuning data. OWASP's LLM02 page also warns that sensitive data a user enters may later appear in model output.
What is the difference between redaction and tokenization?
Redaction removes or masks a value. Tokenization replaces it with a token. Google's Sensitive Data Protection documentation describes pseudonymization as replacing values with cryptographically generated tokens: one-way tokens cannot be reversed, and two-way tokens can be reversed with the key.
How is GenAI DLP different from traditional DLP?
Traditional DLP watches email, files and devices. GenAI DLP adds the AI channel: prompts, uploads and responses. Microsoft's Purview DLP documentation, for example, lists OpenAI ChatGPT and Google Gemini among the AI apps its inline web traffic policies can cover.
Is an open-source PII detector enough on its own?
It can handle detection and anonymization, but you still operate it, tune it and decide what happens to each match. Presidio, for example, describes itself as identification and anonymization modules for private entities in text and images, and says it cannot guarantee it finds all sensitive information.
Can these tools find shadow AI use?
Some browser and network controls can discover which AI apps people use. Microsoft's shadow AI guidance treats discovering apps, blocking unsanctioned ones and blocking sensitive data to sanctioned ones as separate steps. Check each tool's documented discovery scope.
Which OWASP risks do data protection tools cover?
In our scoring they are weighed on LLM02 sensitive information disclosure, LLM07 system prompt leakage and LLM08 vector and embedding weaknesses. The mapping on each tool page is our analysis and cites a documented capability.
