Your intelligence. Your identity. Never connected.
Your intelligence. Your identity. Never connected.
Run powerful AI agents without exposing your payment identity, direct IP address, or unnecessary personal context. Four privacy layers. One isolated architecture.
Built on cryptography. Designed for privacy.
Built on cryptography. Designed for privacy.
Built on cryptography. Designed for privacy.
The Privacy Problem
Powerful AI shouldn’t require a public identity. Every request leaks signals: who paid, where it came from, when it was sent, what the prompt reveals. PARANOID separates them.
4
Independent privacy layers
ZK
Proof-based API payments
Tor
Isolated network routing
Local
Prompt sanitization on-device
Privacy by Design
Every layer protects what the others can’t.
Unlinkable Payments
Zero-knowledge proofs help separate API spending authorization from the on-chain deposit funding it.
Network Isolation
Conversation-specific Tor circuits reduce exposure of your direct IP and help separate network sessions.
Traffic Obfuscation
Parallel decoy requests complicate timing and volume analysis by making real activity less distinguishable.
Local-First Protection
Sensitive context is processed locally before sanitized prompts are transmitted to remote models.
Under the Hood
Privacy engineered into every request.
Pay for intelligence. Not with your identity.
Fund a smart contract vault with ETH or USDC. Groth16 proofs generated locally authorize API spending without revealing which deposit backs each request.
Privacy Payment Console
Anonymous Session
Proof Verified
Groth16
Proof valid
Deposit Linkability
Not disclosed
Ethereum vault
Private API Authorization
example · on-chain settlement
Pay for intelligence. Not with your identity.
Fund a smart contract vault with ETH or USDC. Groth16 proofs generated locally authorize API spending without revealing which deposit backs each request.
Privacy Payment Console
Anonymous Session
Proof Verified
Groth16
Proof valid
Deposit Linkability
Not disclosed
Ethereum vault
Private API Authorization
example · on-chain settlement
Network Isolation
Anonymous Session
Tor Routing
Enabled
Direct IP hidden
Circuit Isolation
Per Chat
isolated
Session Routing · example
Circuit A
Isolated
Circuit B
Isolated
Circuit C
Isolated
Circuit D
Isolated
Circuit E
Isolated
Your requests shouldn’t reveal where you are.
Route agent traffic through isolated Tor circuits. Conversations stay separated at the network layer, and providers see a Tor exit instead of your IP.
Network Isolation
Anonymous Session
Tor Routing
Enabled
Direct IP hidden
Circuit Isolation
Per Chat
isolated
Session Routing · example
Circuit A
Isolated
Circuit B
Isolated
Circuit C
Isolated
Circuit D
Isolated
Circuit E
Isolated
Your requests shouldn’t reveal where you are.
Route agent traffic through isolated Tor circuits. Conversations stay separated at the network layer, and providers see a Tor exit instead of your IP.
Make real activity harder to distinguish.
Controlled decoy requests run alongside real agent traffic, blurring simple usage patterns and raising the cost of timing-based correlation.
Traffic Obfuscation
Anonymous Session
Request Activity
EXAMPLE
Request
Series
Delay
Jitter
1
req-7f3a
Agent
1.2s
↑ 180ms
2
req-c41e
Decoy
0.9s
↑ 95ms
3
req-09bd
Decoy
2.4s
↓ 40ms
4
req-5a72
Agent
1.7s
↑ 210ms
5
req-e18c
Decoy
3.1s
↓ 120ms
6
req-b3f0
Agent
0.8s
↑ 65ms
Make real activity harder to distinguish.
Controlled decoy requests run alongside real agent traffic, blurring simple usage patterns and raising the cost of timing-based correlation.
Traffic Obfuscation
Anonymous Session
Request Activity
EXAMPLE
Request
Series
Delay
Jitter
1
req-7f3a
Agent
1.2s
↑ 180ms
2
req-c41e
Decoy
0.9s
↑ 95ms
3
req-09bd
Decoy
2.4s
↓ 40ms
4
req-5a72
Agent
1.7s
↑ 210ms
5
req-e18c
Decoy
3.1s
↓ 120ms
6
req-b3f0
Agent
0.8s
↑ 65ms
The Stack
Built from independent privacy primitives.
Ethereum
On-chain payment vaults
zkAPI
Private API spending
Groth16
Zero-knowledge proofs
Tor
Network circuit isolation
Cover Traffic
Request pattern obfuscation
Local LLM
Prompt sanitization
Runtime Option
bubblewrap
Sandboxed agent runtime
Runtime Option
MicroVM
Stronger runtime isolation
How it works
Three steps between you and private AI.
1
Fund Privately
Deposit ETH or USDC into the supported Ethereum vault and prepare local zero-knowledge spending credentials.
2
Isolate & Sanitize
Launch the agent inside a sandbox, route its network activity through Tor, and locally sanitize prompts before submission.
3
Execute with Privacy Layers
Submit authorized AI requests with cover traffic while reducing the direct linkage between payments, network origin, and prompt content.
Design Principles
Privacy is an architectural decision.
“Privacy is strongest when independent protections are composed and their limitations are understood.”
DEFENSE IN DEPTH
Security architecture
"
An AI provider should not need to know who funded the request.
ZK PAYMENTS
Payment identity separation
"
A network connection should reveal as little as possible about its origin.
TOR ISOLATION
Network privacy
"
Traffic patterns can reveal information even when message contents are protected.
COVER TRAFFIC
Metadata protection
"
Sensitive context should be minimized before it reaches a remote model.
LOCAL SANITIZATION
Data minimization
"
An AI agent should not inherit unrestricted access to its host environment.
SANDBOX ISOLATION
Runtime containment
Design Principles
Privacy is an architectural decision.
“Privacy is strongest when independent protections are composed and their limitations are understood.”
DEFENSE IN DEPTH
Security architecture
"
An AI provider should not need to know who funded the request.
ZK PAYMENTS
Payment identity separation
"
A network connection should reveal as little as possible about its origin.
TOR ISOLATION
Network privacy
"
Traffic patterns can reveal information even when message contents are protected.
COVER TRAFFIC
Metadata protection
"
Sensitive context should be minimized before it reaches a remote model.
LOCAL SANITIZATION
Data minimization
"
An AI agent should not inherit unrestricted access to its host environment.
SANDBOX ISOLATION
Runtime containment
Design Principles
Privacy is an architectural decision.
“Privacy is strongest when independent protections are composed and their limitations are understood.”
DEFENSE IN DEPTH
Security architecture
"
An AI provider should not need to know who funded the request.
ZK PAYMENTS
Payment identity separation
"
A network connection should reveal as little as possible about its origin.
TOR ISOLATION
Network privacy
"
Traffic patterns can reveal information even when message contents are protected.
COVER TRAFFIC
Metadata protection
"
Sensitive context should be minimized before it reaches a remote model.
LOCAL SANITIZATION
Data minimization
"
An AI agent should not inherit unrestricted access to its host environment.
SANDBOX ISOLATION
Runtime containment
FAQs
Frequently asked questions
How does PARANOID protect my identity?
PARANOID combines independent privacy protections. Zero-knowledge payments reduce the link between deposits and API spending. Tor hides your direct IP from the destination and isolates conversation circuits. Cover traffic makes simple request-pattern analysis harder. Local prompt sanitization reduces sensitive context before prompts leave your device. These layers reduce different forms of exposure, but they do not guarantee complete anonymity.
How do zero-knowledge API payments work?
You deposit ETH or USDC into an Ethereum smart contract vault. The local client generates Groth16 proofs showing valid spending authorization without revealing the specific deposit or note behind it. The payment verifier checks the proof, so API access does not rely on identity-linked billing. Deposits and withdrawals remain subject to the contract’s on-chain rules.
Why use Tor and cover traffic together?
Tor reduces direct exposure of your originating IP address and supports circuit isolation. An observer may still analyze request timing, frequency, and volume. Cover traffic adds decoy activity to complicate simple correlation. Neither mechanism fully defeats a powerful adversary with extensive network visibility.
Can the AI provider still read my prompts?
Yes. A remote model needs access to the prompt it processes. PARANOID uses a local language model to reduce unnecessary identifying details before transmission, but the sanitized prompt remains visible to the remote provider. Local sanitization is risk reduction, not end-to-end confidentiality from the inference provider.
Does PARANOID run AI models locally?
The architecture uses a lightweight local model, such as a compatible Qwen-family model, for prompt sanitization, while larger remote models handle the primary inference task. The agent can run inside a sandbox such as bubblewrap or a microVM. Exact models, sandbox options, and routing capabilities depend on the implementation.
Is PARANOID completely anonymous?
No. PARANOID is designed to reduce identity linkage and metadata exposure, not to eliminate every possible attack. Traffic correlation, sensitive prompt content, endpoint compromise, malicious tools, and implementation errors remain real risks. Stronger protection against sophisticated active adversaries may require mixnets or additional measures.
FAQs
Frequently asked questions
How does PARANOID protect my identity?
PARANOID combines independent privacy protections. Zero-knowledge payments reduce the link between deposits and API spending. Tor hides your direct IP from the destination and isolates conversation circuits. Cover traffic makes simple request-pattern analysis harder. Local prompt sanitization reduces sensitive context before prompts leave your device. These layers reduce different forms of exposure, but they do not guarantee complete anonymity.
How do zero-knowledge API payments work?
You deposit ETH or USDC into an Ethereum smart contract vault. The local client generates Groth16 proofs showing valid spending authorization without revealing the specific deposit or note behind it. The payment verifier checks the proof, so API access does not rely on identity-linked billing. Deposits and withdrawals remain subject to the contract’s on-chain rules.
Why use Tor and cover traffic together?
Tor reduces direct exposure of your originating IP address and supports circuit isolation. An observer may still analyze request timing, frequency, and volume. Cover traffic adds decoy activity to complicate simple correlation. Neither mechanism fully defeats a powerful adversary with extensive network visibility.
Can the AI provider still read my prompts?
Yes. A remote model needs access to the prompt it processes. PARANOID uses a local language model to reduce unnecessary identifying details before transmission, but the sanitized prompt remains visible to the remote provider. Local sanitization is risk reduction, not end-to-end confidentiality from the inference provider.
Does PARANOID run AI models locally?
The architecture uses a lightweight local model, such as a compatible Qwen-family model, for prompt sanitization, while larger remote models handle the primary inference task. The agent can run inside a sandbox such as bubblewrap or a microVM. Exact models, sandbox options, and routing capabilities depend on the implementation.
Is PARANOID completely anonymous?
No. PARANOID is designed to reduce identity linkage and metadata exposure, not to eliminate every possible attack. Traffic correlation, sensitive prompt content, endpoint compromise, malicious tools, and implementation errors remain real risks. Stronger protection against sophisticated active adversaries may require mixnets or additional measures.
FAQs
Frequently asked questions
How does PARANOID protect my identity?
PARANOID combines independent privacy protections. Zero-knowledge payments reduce the link between deposits and API spending. Tor hides your direct IP from the destination and isolates conversation circuits. Cover traffic makes simple request-pattern analysis harder. Local prompt sanitization reduces sensitive context before prompts leave your device. These layers reduce different forms of exposure, but they do not guarantee complete anonymity.
How do zero-knowledge API payments work?
You deposit ETH or USDC into an Ethereum smart contract vault. The local client generates Groth16 proofs showing valid spending authorization without revealing the specific deposit or note behind it. The payment verifier checks the proof, so API access does not rely on identity-linked billing. Deposits and withdrawals remain subject to the contract’s on-chain rules.
Why use Tor and cover traffic together?
Tor reduces direct exposure of your originating IP address and supports circuit isolation. An observer may still analyze request timing, frequency, and volume. Cover traffic adds decoy activity to complicate simple correlation. Neither mechanism fully defeats a powerful adversary with extensive network visibility.
Can the AI provider still read my prompts?
Yes. A remote model needs access to the prompt it processes. PARANOID uses a local language model to reduce unnecessary identifying details before transmission, but the sanitized prompt remains visible to the remote provider. Local sanitization is risk reduction, not end-to-end confidentiality from the inference provider.
Does PARANOID run AI models locally?
The architecture uses a lightweight local model, such as a compatible Qwen-family model, for prompt sanitization, while larger remote models handle the primary inference task. The agent can run inside a sandbox such as bubblewrap or a microVM. Exact models, sandbox options, and routing capabilities depend on the implementation.
Is PARANOID completely anonymous?
No. PARANOID is designed to reduce identity linkage and metadata exposure, not to eliminate every possible attack. Traffic correlation, sensitive prompt content, endpoint compromise, malicious tools, and implementation errors remain real risks. Stronger protection against sophisticated active adversaries may require mixnets or additional measures.
Powerful AI. Minimal exposure.
Explore a privacy-first architecture that separates how AI is paid for, how requests travel, and what personal context reaches the model.
Zero-knowledge payment authorization
Tor isolation and traffic obfuscation
Local prompt sanitization and sandboxing
Powerful AI. Minimal exposure.
Explore a privacy-first architecture that separates how AI is paid for, how requests travel, and what personal context reaches the model.
Zero-knowledge payment authorization
Tor isolation and traffic obfuscation
Local prompt sanitization and sandboxing
Powerful AI. Minimal exposure.
Explore a privacy-first architecture that separates how AI is paid for, how requests travel, and what personal context reaches the model.
Zero-knowledge payment authorization
Tor isolation and traffic obfuscation
Local prompt sanitization and sandboxing