Skip to content

noxtandav/Trishul-AI-Privacy-Framework

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Trishul — The 3-Tier AI Privacy Framework

Named after Shiva's trident (त्रिशूल) — three prongs, three tiers of protection.

A philosophical framework for using AI tools while controlling how much of your personal data you expose to model providers. No custom tools to build — just a mental model and a curated set of existing tools for each tier.


The Problem

When you use AI services like ChatGPT, Claude, or Gemini through their consumer apps, every conversation is tied to your account — your email, your payment info, your identity. Over weeks and months of usage, the provider accumulates a rich profile of who you are: your job, your business ideas, your travel plans, your health concerns, your financial situation, your relationships.

This isn't necessarily malicious. Memory and personalization features are genuinely useful. But the result is that multiple corporations now hold an intimate portrait of your life, and you have limited control over how that data is stored, used, or shared.

Trishul provides a thinking framework for controlling your AI privacy at three distinct levels — from zero precautions to full data isolation — so you can make conscious choices about what you share and with whom.


The Trishul Model

Tier Name What You Protect What You Trade Off
Tier 3 Open Use Nothing Nothing — full convenience
Tier 2 Anonymous Use Your identity and persistent profile Some convenience, slight latency
Tier 1 Full Isolation Everything — identity and content Cost, model capability, or hardware investment

The core insight: not everything you do with AI needs the same level of privacy. Asking for a pasta recipe doesn't need the same protection as analyzing your tax returns. Trishul lets you route each conversation to the appropriate level — like the three prongs of Shiva's trident, each serving a distinct purpose.

Quick Decision Guide

Is this content sensitive at all?
│
├── No → TIER 3 (Open Use)
│         General research, coding, learning, casual chat
│
└── Yes
    │
    ├── Is the CONTENT itself sensitive?
    │   (Financial data, health records, legal docs, trade secrets)
    │
    │   ├── Yes → TIER 1 (Full Isolation)
    │   │         Use Bedrock/Azure or local models
    │   │
    │   └── No, but I don't want it LINKED TO ME
    │             → TIER 2 (Anonymous Use)
    │               Business strategy, competitive research,
    │               personal reflections, sensitive brainstorming
    │
    └── Am I unsure? → Default to TIER 2
        Better to over-protect than under-protect

Tier 3: Open Use

Privacy Level: None Best For: General research, coding help, brainstorming, learning, casual conversation Setup Effort: Zero

Use AI services directly with your account — Claude, ChatGPT, Gemini, Copilot — with all their features enabled. Memory, personalization, conversation history, file uploads, web search. The provider knows who you are and builds a profile over time.

When to Use

  • Looking up technical documentation or getting coding help
  • Brainstorming ideas that aren't competitively sensitive
  • Learning new concepts, getting explanations
  • Writing help for non-sensitive content
  • General research on public topics

What You Get

  • Best model capabilities (all features enabled)
  • Persistent memory across sessions
  • File handling, web browsing, tool use
  • Zero setup or maintenance

What You Accept

  • The provider ties all conversations to your identity
  • A persistent profile is built from your usage patterns
  • Your data is subject to the provider's privacy policy (which can change)
  • Multiple providers may each hold different slices of your personal information

Recommended Tools

No special tools needed. Use the services as-is:

Tool What It Is
Claude Anthropic's assistant — chat, analysis, coding
ChatGPT OpenAI's assistant — chat, DALL-E, browsing
Gemini Google's assistant — deep Google integration
GitHub Copilot AI coding assistant in your IDE
Claude Code Anthropic's CLI coding agent

Just be mindful of what you discuss — this tier works best when you consciously avoid sharing deeply personal, financial, or strategic information.


Tier 2: Anonymous Use (Identity Delinking)

Privacy Level: Medium — content is seen by model providers during inference, but not linked to your identity Best For: Business strategy, competitive research, personal reflections, travel planning, sensitive brainstorming Setup Effort: Minimal — just sign up for OpenRouter and use existing tools

The Core Idea

Instead of building your own chat interface, use existing AI tools that support OpenRouter as a backend. OpenRouter acts as a privacy buffer — your account is with OpenRouter (one entity to trust), and they proxy requests to dozens of model providers without passing your identity downstream.

When an anonymous API call arrives from OpenRouter, the model provider sees a single stateless request — no identity, no account, no email, no link to any previous or future request.

Recommended Tools for Chat

Tool What It Is OpenRouter Support
OpenRouter Chat OpenRouter's built-in chat UI Native
TypingMind Polished chat UI, bring your own API key Yes — set OpenRouter as provider
LibreChat Self-hosted ChatGPT clone Yes — configure as OpenAI-compatible endpoint
Big-AGI Open-source AI chat with multi-model support Yes — OpenAI-compatible
Open WebUI Self-hosted UI for local and remote models Yes — OpenAI-compatible endpoint
Jan Desktop app, local-first with remote model support Yes — OpenAI-compatible

Recommended Tools for Coding (Tier 2)

For multi-step coding tasks where you want identity delinking:

Tool What It Is OpenRouter Support
Aider Terminal-based AI coding assistant Yes — --openai-api-base https://openrouter.ai/api/v1
OpenCode Open-source coding agent Yes — OpenAI-compatible
Cline Autonomous coding agent for VS Code Yes — configure OpenRouter as provider
Continue Open-source AI code assistant for IDEs Yes — OpenAI-compatible
Cursor AI-native code editor Partial — can configure custom API endpoint

Example: Aider with OpenRouter

# Install
pip install aider-chat

# Configure for anonymous use via OpenRouter
export OPENROUTER_API_KEY="your-key-here"
aider --model openrouter/anthropic/claude-sonnet-4-20250514

PII Redaction with Existing Tools

Before sending sensitive content through Tier 2, strip personally identifiable information. You don't need to build a custom redactor — several tools exist:

Tool What It Does Best For
Microsoft Presidio Open-source PII detection and anonymization SDK Programmatic redaction in Python pipelines
PasteGuard Browser extension that redacts PII before pasting Quick manual redaction when copying into chat UIs
Stanza (Stanford NLP) NLP toolkit with NER for 60+ languages Multilingual PII detection
spaCy + custom pipeline Industrial NLP with NER models Building automated redaction into workflows
Amazon Comprehend Managed NLP service with PII detection AWS-native workflows
Google Cloud DLP Cloud-based sensitive data detection GCP-native workflows

Example: Microsoft Presidio

# pip install presidio-analyzer presidio-anonymizer
# python -m spacy download en_core_web_lg

from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine

analyzer = AnalyzerEngine()
anonymizer = AnonymizerEngine()

text = "My name is Rahul Sharma and my email is rahul@sombyte.in"

# Detect PII
results = analyzer.analyze(text=text, language="en")

# Anonymize
anonymized = anonymizer.anonymize(text=text, analyzer_results=results)
print(anonymized.text)
# "My name is <PERSON> and my email is <EMAIL_ADDRESS>"

Example: Manual Approach (No Tools)

For quick, one-off redaction when you don't want to install anything:

  1. Before pasting into any AI chat, mentally scan for: names, emails, phone numbers, company names, addresses, account numbers
  2. Replace them with generic placeholders: "My company" instead of "Acme Corp", "the CEO" instead of "John Smith"
  3. After getting the response, mentally map the placeholders back

This is imperfect but covers the 80% case for casual Tier 2 use.

What PII Redaction Does NOT Catch

  • Contextual identifiers: "I'm the only AI startup founder in my small town who previously worked at [specific company]"
  • Behavioral fingerprints: Unique writing style, timezone patterns, topic combinations
  • Indirect references: "My building on 5th and Main"
  • Domain-specific identifiers: Employee IDs, internal ticket numbers, proprietary terminology

If the content itself is sensitive enough that you wouldn't want anyone to see it, use Tier 1.

Limitations of Tier 2

  • Content is still seen during inference. The model provider processes your tokens. They just can't link them to you.
  • OpenRouter knows your identity. You're trusting one intermediary instead of many providers.
  • PII redaction is imperfect. No tool catches everything. Be mindful of context.
  • Conversation context can leak identity. Highly specific scenarios may be identifying even without explicit PII.

Tier 1: Full Isolation

Privacy Level: Maximum — neither your identity nor your content leaves your trust boundary Best For: Financial documents, tax returns, legal contracts, health records, trade secrets, M&A analysis Setup Effort: Moderate (cloud configuration) to Significant (local hardware)

Option A: Cloud Isolation (AWS Bedrock / Azure AI)

Managed cloud AI services run models within your cloud account boundary. Key guarantees:

  • Your input/output data is not used to train or improve the models
  • The model provider (Anthropic, Meta, etc.) does not have access to your data
  • Data stays within your chosen region
  • Enterprise-grade audit logging (CloudTrail, Azure Monitor)

Recommended tools with Bedrock/Azure:

Tool What It Is Bedrock/Azure Support
Claude Code Anthropic's CLI coding agent Yes — CLAUDE_CODE_USE_BEDROCK=1
Claude Code Desktop (Claude Desktop) Desktop app for Claude Yes — configure Bedrock
Amazon Q Developer AWS's AI coding assistant Native Bedrock
Bedrock Chat AWS Console chat playground Native
Azure AI Foundry Azure's AI studio and playground Native
Open WebUI Self-hosted UI Yes — OpenAI-compatible with Bedrock gateway

Claude Code with Bedrock (the simplest Tier 1 setup for coding):

# Set environment variables
export CLAUDE_CODE_USE_BEDROCK=1
export AWS_PROFILE=your-profile
export AWS_REGION=us-east-1
export ANTHROPIC_MODEL=us.anthropic.claude-sonnet-4-20250514-v1:0

# Use Claude Code normally — all data stays in your AWS account
claude

Additional hardening:

  • VPC Endpoint for Bedrock (no internet transit)
  • CloudTrail logging for full audit trail
  • KMS encryption with your own keys
  • IAM policies to restrict who can invoke models

Option B: Local Models (Full Air-Gap)

For absolute isolation, run open-source models on your own hardware. Nothing leaves your machine.

Recommended tools:

Tool What It Is
Ollama Simplest way to run local models — one command setup
LM Studio Desktop app for running local models with a chat UI
Jan Desktop app, local-first design
Open WebUI Self-hosted chat UI that connects to Ollama
LocalAI OpenAI-compatible local API server
llama.cpp Bare-metal inference engine for maximum performance
Aider Coding assistant — works with Ollama local models

Quick start with Ollama:

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh    # Linux/macOS
# Or download from https://ollama.com for Windows

# Pull a capable model
ollama pull llama3.1:70b      # Best quality, needs ~40GB RAM
ollama pull llama3.1:8b       # Lighter, needs ~8GB RAM
ollama pull qwen2.5:32b       # Good for structured analysis
ollama pull deepseek-r1:14b   # Good for reasoning tasks

# Start chatting
ollama run llama3.1:8b

Hardware requirements:

Model Size RAM Required GPU VRAM Quality Level
7-8B params 8 GB 8 GB Good for simple tasks, summaries
13-14B params 16 GB 16 GB Decent for analysis, writing
32B params 24 GB 24 GB Good for complex reasoning
70B params 40-48 GB 2x 24 GB Near-frontier for most tasks

Choosing Between Cloud and Local

Factor Bedrock / Azure Local Models
Trust boundary Your cloud account (with legal guarantees) Your physical hardware
Model quality Frontier models (Claude, GPT-4 class) Open-source (Llama, Qwen, DeepSeek)
Setup effort AWS/Azure account + IAM configuration Hardware purchase + software setup
Running cost Per-token pricing (~$3-15 per million tokens) Electricity + hardware amortization
Scalability Unlimited Limited by your hardware
Best for Complex analysis needing frontier intelligence Simpler tasks, absolute zero-trust

Limitations of Tier 1

  • Cloud option: You're trusting AWS/Azure as infrastructure providers. They process your data on their servers, even if the model provider can't see it.
  • Local option: Model quality is significantly lower than frontier models. Open-source models are improving rapidly but still lag behind Claude/GPT-4 for complex reasoning.
  • Cost: Bedrock/Azure is more expensive per token than direct API. Local hardware requires $500-2000+ upfront.
  • Maintenance: Both options require ongoing infrastructure management.

Choosing the Right Tier — Examples

Use Case Recommended Tier Why
"How do I write a for loop in Python?" Tier 3 Zero sensitivity
"What's the best restaurant in Tokyo?" Tier 3 No personal data
"Help me plan a marketing strategy for my SaaS" Tier 2 Competitive strategy, don't want it profiled
"Analyze my competitors and their pricing" Tier 2 Business intelligence, identity-sensitive
"I'm feeling stressed about my startup failing" Tier 2 Personal emotional content
"Review my tax return and suggest deductions" Tier 1 Financial documents with PII
"Analyze this bank statement for spending patterns" Tier 1 Sensitive financial data
"Summarize this legal contract for my company" Tier 1 Confidential business documents
"What medications interact with my prescription?" Tier 1 Health information

Threat Model Summary

Threat Tier 3 Tier 2 Tier 1 (Cloud) Tier 1 (Local)
Provider builds a persistent profile of you Exposed Protected Protected Protected
Provider links your conversations across sessions Exposed Protected Protected Protected
Provider uses your data for training Exposed Varies Contractually protected Protected
Provider sees your content during inference Exposed Exposed AWS/Azure processes it Protected
Third party intercepts data in transit TLS only TLS only VPC endpoint available Air-gapped
Government subpoena to provider Exposed OpenRouter only AWS/Azure only Protected
Data breach at provider Exposed Content exposed Content exposed Protected

Trust Boundaries

Tier 3:  You ──→ Provider (full trust in provider)
Tier 2:  You ──→ OpenRouter ──→ Provider (trust OpenRouter with identity,
                                           provider sees anonymous content)
Tier 1a: You ──→ AWS/Azure (trust cloud infra, not model provider)
Tier 1b: You ──→ Your hardware (trust no one)

The Practical Setup (Three Shell Aliases)

The simplest implementation of Trishul is three aliases in your shell config:

# ~/.bashrc or ~/.zshrc

# Tier 3: Open use — Claude Code with your Anthropic account
alias ai="claude"

# Tier 2: Anonymous use — Aider via OpenRouter
alias ai-anon="OPENAI_API_BASE=https://openrouter.ai/api/v1 OPENAI_API_KEY=$OPENROUTER_API_KEY aider --model openrouter/anthropic/claude-sonnet-4-20250514"

# Tier 1: Full isolation — Claude Code via AWS Bedrock
alias ai-secure="CLAUDE_CODE_USE_BEDROCK=1 AWS_PROFILE=bedrock AWS_REGION=us-east-1 claude"

That's it. Three commands, three privacy levels. No backend, no frontend, no Docker.


FAQ

Q: Do I need to build anything? A: No. Trishul is a thinking framework. The tools already exist — you just need to choose the right ones for each tier and be intentional about which tier you use for each conversation.

Q: Can I use a VPN with Tier 2 for extra anonymity? A: If you're using OpenRouter's API directly (from a server or CLI tool), the model provider never sees your IP — only OpenRouter's. A VPN would only mask your IP from OpenRouter itself.

Q: Is OpenRouter trustworthy? A: OpenRouter is a single point of trust in Tier 2. Review their privacy policy. The tradeoff is trusting one company instead of five model providers separately.

Q: How much does each tier cost? A: Tier 3 is whatever your existing subscriptions cost ($0-20/month). Tier 2 adds OpenRouter API costs (typically $1-10/month for moderate use). Tier 1 with Bedrock adds 2-5x the per-token cost. Tier 1 local requires $500-2000+ in GPU hardware upfront.

Q: What if I accidentally use the wrong tier? A: Build the habit of pausing before each conversation: "Is this sensitive? Is the content itself sensitive, or just the identity link?" The decision flowchart above becomes second nature after a week.

Q: Can PII redaction tools handle languages other than English? A: Microsoft Presidio supports multiple languages. Stanza supports 60+ languages. For manual redaction, the approach is language-agnostic — just replace identifying terms with placeholders before pasting.

Q: What about file uploads (PDFs, images)? A: For Tier 2, you'd need to redact PII from files before uploading — significantly harder than text. For documents with embedded PII (financial statements, ID scans), use Tier 1 directly.

Q: Daniel Miessler's PAI/Kai system doesn't address privacy at all. Why should I care? A: Daniel's system is optimized for maximum productivity with full trust in Anthropic (Tier 3). That's a valid choice for many use cases. Trishul exists for people who want to be intentional about which conversations get that level of trust and which don't.


Documentation

Contributing

Contributions are welcome! See CONTRIBUTING.md for details.

Areas we'd especially love help with:

  • Tool recommendations and reviews for each tier
  • PII redaction tool comparisons and benchmarks
  • Guides for setting up Bedrock/Azure with specific tools
  • Translations of the framework into other languages
  • Real-world usage patterns and case studies

License

Released under the MIT License.


Trishul (त्रिशूल) — because privacy is not binary. It's a spectrum, and you should be able to choose where you sit on it for each interaction.

About

A 3-tier framework for controlling your AI privacy — from open use to full isolation.

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors