Add PII redaction

Author: danbarr

docs/about/changelog.md

@@ -13,6 +13,11 @@ Major features and changes are noted here. To review all updates, see the
 
 Related: [Upgrade CodeGate](../how-to/install.md#upgrade-codegate)
 
+- **PII redaction:** - 10 Feb, 2025\
+  Starting with v0.1.18, CodeGate now redacts personally identifiable
+  information (PII) found in LLM prompts and context. See the
+  [feature page](../features/secrets-encryption.md) to learn more.
+
 - **Model muxing** - 7 Feb, 2025\
   With CodeGate v0.1.17 you can use the new `/v1/mux` endpoint to configure
   model selection based on your workspace! Learn more in the

docs/features/dependency-risk.md

@@ -1,15 +1,14 @@
 ---
 title: Dependency risk awareness
 description: Protection from malicious or vulnerable dependencies
-sidebar_position: 20
 ---
 
 ## What's the risk?
 
 The large language models (LLMs) that drive AI coding assistants are incredibly
 costly and time-consuming to train. That's why each one has a "knowledge cutoff
 date" which is often months or even years in the past. For example, GPT-4o's
-training cutoff was October 2023\.
+training cutoff was October 2023.
 
 But the open source software ecosystem moves quickly, and so do malicious actors
 seeking to exploit the software supply chain. LLMs often suggest outdated,

docs/features/muxing.md

@@ -1,7 +1,6 @@
 ---
 title: Model muxing
 description: Configure a per-workspace LLM
-sidebar_position: 35
 ---
 
 ## Overview

docs/features/secrets-encryption.md

@@ -1,62 +1,90 @@
 ---
-title: Secrets encryption
+title: Secrets encryption and PII redaction
 description: Keep your secrets a secret
-sidebar_position: 10
 ---
 
 ## What's the risk?
 
-As you interact with an AI coding assistant, sensitive data like passwords and
-access tokens can be unintentionally exposed to third-party providers through
-the code snippets and files you share as context. These secrets may become part
-of the training data used to improve the AI model and potentially be exposed to
-other users.
+As you interact with an AI coding assistant, sensitive data like passwords
+access tokens, and even personally identifiable information (PII) can be
+unintentionally exposed to third-party providers through the code and files you
+share as context. Besides the privacy and regulatory implications of exposing
+this information, it may become part of the AI model's training data and
+potentially be exposed to future users.
 
 ## How CodeGate helps
 
 CodeGate helps you protect sensitive information from being accidentally exposed
 to AI models and third-party AI provider systems by redacting detected secrets
-from your prompts using encryption.
+and PII found in your prompts.
 
 ## How it works
 
-CodeGate automatically scans all prompts for secrets such as:
+CodeGate automatically scans all prompts for secrets and PII. This happens
+transparently without requiring a specific prompt. Without interrupting your
+development flow, CodeGate protects your data by encrypting secrets and
+anonymizing PII. These changes are made before the prompt is sent to the LLM and
+are restored when the result is returned to your machine.
 
-- API keys and tokens
-- Private keys and certificates
-- Database credentials
-- SSH keys
-- Cloud provider credentials
-
-This scan happens transparently without requiring a specific prompt.
+When a secret or PII is detected, CodeGate adds a message to the LLM's output
+and an alert is recorded in the [dashboard](../how-to/dashboard.md) (PII alerts
+in the dashboard are coming soon).
 
 :::info
 
 Since CodeGate runs locally, your secrets never leave your system unprotected.
 
 :::
 
-CodeGate transparently encrypts secrets before sending the prompt to the LLM.
-This way, CodeGate protects your sensitive data without blocking your
-development flow. This is performed on the fly using AES256-GCM encryption with
-a temporary per-session key that is securely erased from memory after the
-response is delivered to your plugin.
-
 ```mermaid
 sequenceDiagram
     participant Client as AI coding<br>assistant
     participant CodeGate as CodeGate<br>(local)
     participant LLM as AI model<br>(remote)
 
-    Client ->> CodeGate: Prompt with<br>plaintext secrets
+    Client ->> CodeGate: Prompt with<br>plaintext secrets/PII
     activate CodeGate
-    CodeGate ->> LLM: Prompt with<br>encrypted secrets
+    CodeGate ->> LLM: Prompt with<br>redacted secrets/PII
     deactivate CodeGate
     activate LLM
-    note right of LLM: LLM only sees<br>encrypted values
-    LLM -->> CodeGate: Response with<br>encrypted secrets
+    note right of LLM: LLM only sees<br>redacted values
+    LLM -->> CodeGate: Response with<br>redacted data
     deactivate LLM
     activate CodeGate
-    CodeGate -->> Client: Response with<br>plaintext secrets
+    CodeGate -->> Client: Response with<br>original data
     deactivate CodeGate
 ```
+
+### Secrets encryption
+
+CodeGate uses pattern matching to detect secrets such as:
+
+- API keys and tokens
+- Private keys and certificates
+- Database credentials
+- SSH keys
+- Cloud provider credentials
+- ...and more - see the
+  [signatures file](https://github.com/stacklok/codegate/blob/main/signatures.yaml)
+  in the project repo
+
+CodeGate transparently encrypts secrets before sending the prompt to the LLM.
+This is performed on the fly using AES256-GCM encryption with a temporary
+per-session key. When the LLM returns a response, CodeGate decrypts the secret
+before delivering it to your coding assistant, then securely erases the
+temporary key from memory.
+
+### PII redaction
+
+CodeGate scans for common types of PII like:
+
+- Email addresses
+- Phone numbers
+- Government identification numbers
+- Credit card numbers
+- Bank accounts and crypto wallet IDs
+
+CodeGate anonymizes PII by replacing each string with a unique identifier before
+sending the prompt to the LLM. This way, CodeGate protects your sensitive data
+without blocking your development flow. When the LLM returns a response,
+CodeGate matches up the identifier and replaces it with the original value.

docs/features/security-reviews.md

@@ -1,7 +1,6 @@
 ---
 title: Security reviews
 description: Enhanced secure coding guidance
-sidebar_position: 30
 ---
 
 ## What's the risk?

docs/features/workspaces.mdx

@@ -1,7 +1,6 @@
 ---
 title: Workspaces
 description: Organize and customize your project environments
-sidebar_position: 40
 ---
 
 import useBaseUrl from '@docusaurus/useBaseUrl';

docs/index.md

@@ -36,8 +36,9 @@ sequenceDiagram
 CodeGate includes several key features for privacy, security, and coding
 efficiency, including:
 
-- [Secrets encryption](./features/secrets-encryption.md) to protect your
-  sensitive credentials
+- [Secrets encryption and PII redaction](./features/secrets-encryption.md) to
+  protect your sensitive credentials and anonymize personally identifiable
+  information
 - [Dependency risk awareness](./features/dependency-risk.md) to update the LLM's
   knowledge of malicious or deprecated open source packages
 - [Model muxing](./features/muxing.md) to quickly select the best LLM
@@ -101,7 +102,7 @@ Review the [installation instructions](./how-to/install.md).
 
 Learn more about CodeGate's features:
 
-- [Secrets encryption](./features/secrets-encryption.md)
+- [Secrets and PII redaction](./features/secrets-encryption.md)
 - [Dependency risk awareness](./features/dependency-risk.md)
 - [Security reviews](./features/security-reviews.md)
 - [Workspaces](./features/workspaces.mdx)