Model Armor is a managed Google Cloud service that screens prompts and responses of large language model applications. It detects prompt injection and jailbreaks, filters harmful content such as hate speech, sexually explicit and dangerous content, recognises sensitive data through Sensitive Data Protection and scans for malicious URLs. Documents are screened as well. Administrators set templates and confidence thresholds, choose between inspect-only and inspect-and-block, and call the service through an API. According to Google, Model Armor can also protect AI applications running on other clouds.
Model Armor is a product of Google LLC, part of Alphabet Inc., headquartered in the United States. US legislation such as the CLOUD Act may therefore apply to what the service processes. According to the documentation, processing is stateless and happens in memory, regional endpoints exist, and data residency controls keep processing within a chosen geography such as the US or the EU. Google's general Cloud Data Processing Addendum names Google as processor and defines the GDPR.