liveBook · Manning

Sanitization

As explained in AI Governance

Input sanitization is the modification of supplied data to remove or neutralize content that could violate security or processing requirements.

Sanitization modifies data to reduce security, privacy, or processing risks. In an organization-controlled LLM integration, a pre-call pipeline can validate payloads, anonymize sensitive data, and encrypt protected fields before sending a request to a vendor. section 2.3.1

Apply sanitization at relevant processing boundaries:

  • Before indexing documents, strip hidden formatting and metadata that could conceal malicious instructions. section 4.6.3
  • After an LLM call, remove sensitive or regulated information and format outputs according to internal standards. section 2.3.1

Sanitization reduces exposure but cannot guarantee that a model will treat retrieved untrusted text as reference material. Limit risky sources, make citations visible, and monitor for suspicious content and behavior. section 4.6.3 section 2.3.1

How AI Governance covers it

How deep it goes
Shows sanitization before API calls, after calls, and before indexing, but specifies no algorithms or guarantees of prevention.
Examples
Uses payload validation, sensitive-data anonymization, protected-field encryption, output removal, and stripping hidden document formatting and metadata.
What you can do after
Operate/configure pre-call filters to validate, anonymize, or encrypt payloads and strip hidden formatting and metadata before indexing.
How it connects
Connects input validation, anonymization, encryption, output filtering, document poisoning, retrieval, and audit logging.

Also explained in

Compare how 4 books explain sanitization →

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.