AI Model Providers and Data Handling

Which providers run Guard's AI models, what each one does, and the contract and service terms that keep them from training on your data.

Overview

Praetorian Guard Platform (PGP) runs AI models through three providers: Amazon Bedrock, Fireworks.ai, and Typesafe.ai. PGP sends every model request through its own AI gateway, which decides which provider serves it. None of these providers trains models on your data. We hold a contract and terms of service with each provider, and those terms rule out training on the prompts and outputs PGP sends.

This article covers who hosts the models and how each provider handles request data. To choose which model vendors your account may use, see AI Model Providers. For how PGP isolates and anonymizes your data before any model sees it, see AI Data Privacy and Tenant Isolation.

Providers at a Glance

Provider	What it runs for PGP	Trains on your data?
Amazon Bedrock	Frontier and open-source foundation models	No
Fireworks.ai	Praetorian proprietary models	No
Typesafe.ai	Typesafe decision models	No

Amazon Bedrock

Amazon Bedrock hosts the frontier and open-source foundation models that PGP uses, such as Anthropic Claude, Meta Llama, Mistral, DeepSeek, and Amazon Nova.

No training: Bedrock does not use prompts or completions to train or improve foundation models, and it does not distribute them to other parties.
No model-provider access: AWS runs each model in accounts that the Bedrock service team owns. The company that built the model cannot access those accounts, so it never sees your prompts, completions, or Bedrock logs.
Retention: Any request data AWS keeps for abuse detection stays inside AWS. It is not shared with the model provider.
Network path: PGP calls Bedrock through a private interface VPC endpoint, so this traffic stays on the AWS network.
Access control and audit: PGP authenticates to Bedrock with an IAM task role scoped to Bedrock model access, not with long-lived keys, and AWS CloudTrail logs Bedrock API activity.
Region: Some models are called through Bedrock cross-region inference profiles, which can process a request in another AWS Region in the same geography as your PGP deployment.

Fireworks.ai

Fireworks.ai hosts Praetorian's proprietary models.

No training: Our contract and the Fireworks terms bar Fireworks from using prompts, completions, or other request content to train or improve any model.
No persistent storage of content: Prompts and completions stay in memory only for the life of the request and are never written to persistent storage. Fireworks logs request metadata, such as token counts, to run the service. PGP doesn't use the Fireworks features that store content, such as stored responses or opt-in logging for tuning.
Network path: PGP's AI gateway calls Fireworks over HTTPS from PGP's private subnets through a NAT gateway. The Fireworks API key is stored encrypted in AWS Systems Manager Parameter Store.

Typesafe.ai

Typesafe.ai provides decision models that PGP uses for classification and ranking.

No training: Typesafe's terms say it will not train or fine-tune any AI or machine learning model on customer prompts or other input.
Data processing: Typesafe processes request data under our agreement and its Data Processing Addendum, which limit use of that data to our documented instructions. Typesafe does not sell customer data.
Network path: Requests to Typesafe are encrypted in transit over HTTPS.

What Applies to Every Provider

Anonymization first: Before findings reach the AI knowledge system, PGP replaces sensitive identifiers such as credentials, email addresses, and IP addresses with typed placeholders. See AI Data Privacy and Tenant Isolation.
Tenant scope: A model request carries only data from the account that made it.
Encrypted in transit: Every connection between PGP and a model provider is encrypted with TLS.
Vendor choice: A super-admin or a user with manage settings can limit which model vendors PGP may use. See AI Model Providers.