OpenAI privacy-filter for PII detection and masking is now available in Amazon SageMaker JumpStart
Deploy OpenAI's PII detection model on your own AWS infrastructure in minutes—keeping sensitive data out of your AI pipelines without leaving your account.
View original announcement →Visual Summary
graph TD
A{{OpenAI Privacy-Filter on JumpStart}}:::announced
B(Amazon SageMaker JumpStart):::compute
C(SageMaker Endpoint):::compute
D([PII Detection & Masking]):::feature
E([Tunable Thresholds]):::feature
F((Data Engineers)):::external
G(SageMaker Studio):::compute
H(SageMaker Python SDK):::compute
I([8 PII Categories]):::feature
F ==>|"deploys model"| G
F -->|"programmatic deploy"| H
G ==>|"few clicks"| B
H -->|"SDK call"| B
B ==>|"hosts"| A
A ==>|"runs on"| C
C -->|"performs"| D
D -->|"configurable"| E
D -->|"detects"| I
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
AWS has made OpenAI's privacy-filter model available in Amazon SageMaker JumpStart, enabling customers to deploy a purpose-built PII detection and masking solution directly within their AWS environment. The model is a bidirectional token-classification model capable of identifying and labeling sensitive data spans—such as names, addresses, phone numbers, emails, account numbers, URLs, dates, and secrets—in a single forward pass. Customers can deploy it with just a few clicks via SageMaker Studio or programmatically through the SageMaker Python SDK.
How It Works
- Bidirectional token classification: The model processes input text in both directions simultaneously, giving it full context around each token before assigning a PII label, which improves accuracy over unidirectional approaches.
- Single forward pass inference: The entire input sequence is labeled in one pass through the network, making inference fast and computationally efficient for high-throughput pipelines.
- PII span detection: The model identifies and tags eight categories of sensitive data: account numbers, physical addresses, email addresses, personal names, phone numbers, URLs, dates, and secrets (e.g., API keys, passwords).
- Tunable thresholds: Confidence thresholds and detection sensitivity can be adjusted, allowing teams to balance precision and recall based on their specific compliance or risk tolerance requirements.
- SageMaker JumpStart deployment: The model is hosted in the SageMaker JumpStart model catalog and can be deployed to a managed SageMaker endpoint via the Studio UI or the
sagemakerPython SDK with minimal configuration. - AWS-managed infrastructure: Once deployed, the model runs on AWS compute (e.g., EC2-backed SageMaker instances), meaning data never needs to leave the customer's AWS account or VPC boundary.
Why It's Important
- Data sanitization at scale: Organizations processing large volumes of text—logs, support tickets, documents, chat transcripts—can now automate PII scrubbing before feeding data into downstream ML pipelines or storage systems, reducing compliance risk.
- Privacy-by-design for AI workflows: Teams building RAG pipelines, fine-tuning datasets, or analytics systems can integrate PII masking as a preprocessing step, preventing sensitive data from inadvertently entering foundation model prompts or training sets.
- Regulatory compliance enablement: Industries subject to GDPR, HIPAA, CCPA, and similar regulations gain a readily deployable tool to demonstrate data minimization and pseudonymization controls without building custom NLP solutions.
- Reduced time-to-value: Availability in JumpStart eliminates the need to source, containerize, and operationalize the model independently—cutting deployment effort from days to minutes.
- OpenAI model on AWS infrastructure: This expands the cross-cloud model ecosystem, giving AWS customers access to an OpenAI-developed model while keeping data and compute entirely within their AWS environment.
How It's Different
- Purpose-built for PII, not general NLP: Unlike general-purpose LLMs that can be prompted to find PII, privacy-filter is a specialized token-classification model optimized specifically for this task, offering lower latency and more predictable behavior.
- Single forward pass vs. generative inference: Generative models require autoregressive decoding, which is slower and more expensive; privacy-filter labels the full sequence in one pass, making it significantly faster and cheaper per token.
- Tunable sensitivity: Compared to rule-based regex systems (e.g., AWS Comprehend's built-in PII detection), the model is context-aware and can catch semantically implied PII that pattern matching would miss, while still allowing threshold tuning.
- On-premises design philosophy brought to cloud: The model was originally designed for on-premises high-throughput workflows, meaning it is optimized for batch and streaming efficiency—a characteristic that translates well to SageMaker batch transform jobs.
- OpenAI provenance in a non-OpenAI cloud: Unlike using OpenAI's API directly, this deployment keeps all data within AWS, addressing data residency and egress concerns that prevent some enterprises from using OpenAI's hosted endpoints.
- JumpStart catalog integration: Unlike deploying a Hugging Face model manually, JumpStart provides a curated, tested deployment path with pre-configured container images and IAM integration out of the box.
When to Prefer It
- Pre-processing pipelines for LLM fine-tuning: When building training datasets from raw enterprise data, use privacy-filter to scrub PII before the data is used to fine-tune or instruct-tune a foundation model.
- Real-time API gateway sanitization: When building APIs that accept user-submitted text and forward it to an LLM, deploy privacy-filter as an inline Lambda or SageMaker endpoint to mask PII before the prompt is sent downstream.
- Compliance-driven data archival: When ingesting documents, emails, or records into a data lake for analytics, use privacy-filter in a SageMaker batch transform job to produce sanitized copies for broader team access.
- High-throughput log anonymization: When processing application or system logs at scale that may contain user identifiers, addresses, or credentials, the model's speed and single-pass design make it suitable for streaming pipelines via Kinesis or Kafka integration.
- Regulated industry AI adoption: When operating in healthcare, finance, or legal sectors where demonstrating PII controls is a prerequisite for deploying AI tools, privacy-filter provides an auditable, model-based sanitization layer.
- Teams without NLP expertise: When a team lacks the resources to train or fine-tune a custom NER model for PII detection, JumpStart's one-click deployment provides an immediately usable, production-grade solution.
Availability
- GA status: The model is generally available in Amazon SageMaker JumpStart as of July 13, 2026.
- Access method: Available via the Models landing page in Amazon SageMaker Studio or programmatically through the SageMaker Python SDK.
- Regional availability: Specific supported regions are not enumerated in the announcement; customers should check the SageMaker JumpStart model catalog within their target region to confirm availability.
- Pricing model: No separate model licensing fee is mentioned; customers pay standard SageMaker endpoint instance costs (per-hour compute) for the underlying ML instance used to host the model.
- License responsibility: As with all JumpStart third-party models, customers are responsible for reviewing and complying with the applicable license terms from OpenAI before deploying or using the model.
- Limitation — on-premises note: The announcement describes the model as designed for on-premises workflows; when deployed on SageMaker, teams should validate that latency and throughput meet their requirements for their specific instance type selection.
