AI chat service Obscurify.ai has released an open-source library designed to prevent AI providers from linking separate conversations to the same user. The technology, called unlinkable inference, uses blind-signed digital vouchers and prepaid sessions to verify payment without attaching a persistent account identifier to each AI request. Obscurify says its web chat and Android app now use the system, while the library is available for other AI services to integrate into their own infrastructure. The release adds a new privacy-focused approach to an industry where account-based authentication, usage tracking, and conversation histories can make supposedly private AI interactions traceable.
Obscurify Open-Sources Privacy Technology for AI Chats
The release addresses a structural privacy issue in AI services: even when users do not provide their names or email addresses, persistent account identifiers can allow providers to associate separate conversations with the same person. This creates a potential exposure risk when a user shares identifying information in one interaction. If conversations are linked through a common account key, that information may make other interactions attributable to the same user.
Obscurify’s approach separates payment verification from conversation authentication. Instead of attaching an account identifier to each request, the system uses digital vouchers that the provider can validate without knowing which account originally received them.
How blind signatures make conversations harder to link
The mechanism is based on blind signatures, a cryptographic technique introduced by David Chaum in the early 1980s and used in privacy-oriented digital cash systems.
In Obscurify’s implementation, the user’s device maintains a wallet of prepaid digital vouchers. During the issuance process, the vouchers are blinded before being signed by the service. The provider can charge the user’s account and sign the blinded vouchers, but it cannot see their underlying contents.
When a conversation begins, the client submits a small number of vouchers to open a prepaid session. The session is identified by a randomly generated secret rather than the user’s account key. Subsequent messages use that session secret for authentication, while the provider verifies the vouchers to establish that the session has been paid for.
After five minutes of inactivity, the session closes and its unused balance is returned to the device as newly signed vouchers. According to Obscurify, the returned vouchers have not previously been seen by the service, helping prevent the provider from connecting the original payment with the conversation.
The company says users pay for the tokens they actually consume, with billing calculated to one-thousandth of a cent. Voucher wallets and session balances can also be reconstructed using the user’s 16-digit account number, allowing recovery after a device reset or loss of local browser data.
Open-source library targets AI developers
The release extends beyond Obscurify’s own applications. Its Rust-based unlinkable-inference library includes client-side and server-side components, allowing AI providers to implement the mechanism within their existing infrastructure rather than relying on an external privacy broker.
The project provides several integration options:
- Rust library: Core functionality for voucher wallets, session management, and server-side voucher verification, under MIT or Apache 2.0 licensing.
- WebAssembly package: An npm distribution that enables web applications to use the client-side functionality.
- Command-line proxy: A GPL-licensed tool designed to support OpenAI- and Anthropic-compatible APIs, including coding agents.
- Technical documentation: The repository includes a wire protocol specification and threat model, giving developers material to evaluate the design and its security assumptions.
The proxy is particularly relevant as developers increasingly use AI coding agents and other tools that connect to model providers through standard APIs. Obscurify says the proxy can manage private sessions without requiring changes to the underlying agent workflow. The regular API key is used to replenish the voucher wallet, while individual conversations are authenticated using session secrets.
Obscurify has enabled unlinkable inference by default for signed-in users of its web chat and Android app, without an additional charge. Subscribers can also receive monthly anonymous passes that reveal their subscription tier, allowing sessions to access the models included in their plan without exposing the account identity.
Privacy benefits come with important limitations
Unlinkable inference does not make AI conversations completely anonymous or prevent the service from accessing their content. Obscurify’s documentation explicitly states that the provider can still read user prompts because it must forward them to the selected AI model.
The service can also see a user’s IP address unless the connection is routed through Tor or a VPN. Timing information may create additional correlation risks, particularly when relatively few people are using the system at a given time. A prompt containing a name, workplace, address, or other identifying detail can therefore still reveal the user’s identity.
These limitations matter for organisations considering the technology for sensitive workflows. Removing persistent account identifiers can reduce one avenue for linking conversations, but it does not replace prompt minimisation, network privacy controls, access governance, or appropriate data-handling policies.
There is also an important standards distinction. RSA blind signatures are specified in RFC 9474, published in October 2023 through the Internet Research Task Force’s Crypto Forum Research Group. The RFC is informational; it is not an Internet Standards Track specification or a formal IETF standard.
A new privacy option for the AI infrastructure stack
The release arrives as AI providers face growing scrutiny over how they authenticate users, retain conversation histories, and handle sensitive prompts. Privacy controls have traditionally focused on encryption, data retention, and whether conversations are used to train models. Obscurify’s approach addresses a different layer: whether separate requests can be associated with the same account in the first place.
If adopted more broadly, unlinkable inference could give AI platforms another way to reduce the amount of identity information attached to routine model interactions. Its open-source availability may also make experimentation easier for providers that want to deploy the mechanism within their own systems.
However, adoption will depend on implementation quality, security review, compatibility with existing billing systems, and confidence in the cryptographic design. The library’s publication makes the architecture available for inspection, but it does not by itself establish independent security validation or widespread production adoption.
For now, Obscurify has introduced a technically distinct approach to AI privacy: verifying that a conversation is paid for without requiring the service to know which account funded it. Whether that model becomes a broader infrastructure pattern will depend on developer uptake and how effectively its protections hold up under independent scrutiny.
Market Landscape
AI privacy is increasingly becoming an infrastructure issue rather than simply a matter of account settings or data-retention policies. Providers must balance user authentication, billing, abuse prevention, model access, and operational monitoring while limiting the personal information associated with individual requests.
Obscurify’s release sits within several related technology trends:
- Privacy-preserving payments: Blind signatures and digital cash techniques allow services to validate payment credentials without directly linking each transaction to a user’s identity.
- Private AI authentication: Session-based credentials offer an alternative to persistent API keys and account identifiers for individual conversations.
- Open-source privacy infrastructure: Reusable libraries can help AI providers integrate privacy mechanisms into their own applications and backend systems.
- Agentic AI security: Coding agents and API-connected applications create additional demand for authentication methods that limit the exposure of long-lived credentials.
The technology does not replace encryption or guarantee that prompts cannot be identified. Instead, it targets the relationship between account identity and conversation activity. Its strongest potential differentiator is the ability to separate payment records from individual AI sessions without introducing a third-party broker.
Top Insights
- Obscurify’s open-source library uses blind-signed vouchers to authenticate prepaid AI sessions without attaching persistent account identifiers to individual conversations.
- The Rust library includes server-side and client-side components, with a WebAssembly package for browser integration and a command-line proxy for AI agents.
- Obscurify has enabled the technology by default in its web chat and Android app for signed-in users, without additional charges.
- Prompts and IP addresses remain visible to the service, meaning unlinkable inference does not provide complete anonymity or end-to-end encryption.
- RFC 9474 specifies RSA blind signatures as an informational IRTF document, rather than a formal Internet Standards Track specification.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
