Yanez SN54

The internet has spent thirty years getting better at recognising accounts. It remains surprisingly bad at recognising people.

For much of the internet’s history, that distinction was manageable. An email address, password and perhaps a phone number were usually enough to establish that someone belonged behind an account. Artificial intelligence makes this assumption increasingly fragile. Bots can maintain convincing conversations, deepfakes can imitate faces and voices, synthetic identities can be assembled from stolen or generated information, and AI agents can perform actions at a speed and scale that no human operator could match.

The question facing banks, payment networks, online communities and governments is therefore changing. They no longer need to know only whether an account possesses the correct credentials. They increasingly need to establish whether a real human exists behind an action, whether that person is unique and, in the case of an AI agent, whether the agent is acting with legitimate human authority.

Yanez, Bittensor Subnet 54, is building infrastructure around that problem. Its broader ambition is to provide privacy-preserving proof of humanhood and proof of uniqueness: evidence that someone is a real and distinct person without requiring that person to reveal their full identity or surrender biometric data to another central database.

The problem with proving that someone is human

Traditional identity systems usually begin by collecting information. A bank asks for a passport. A platform requests a phone number. A biometric provider records a face, fingerprint or voice sample. The information is then checked against documents, databases or models controlled by a trusted institution.

There are legitimate reasons for this approach. Banks still need to meet legal identification requirements, and proof of humanhood cannot replace every form of KYC. Yet collecting more identity data every time trust is required creates its own risks. Sensitive information becomes concentrated inside databases that must be protected indefinitely, while users are forced to trust institutions they may barely know with documents and biometrics that cannot simply be changed after a breach. A compromised password can be replaced. A face is slightly less cooperative.

The problem becomes particularly awkward in systems that do not need to know someone’s legal identity. A DAO may only need to establish that each participant is one unique human. An airdrop may need to distinguish genuine users from a farm of automated accounts. An online platform may want to reduce bot activity without asking every visitor to upload a passport.

In these situations, the verifier does not necessarily need the underlying identity data. It needs a reliable answer to a narrower question.

Yanez is designed around that distinction.

According to the team, its system can produce cryptographically verifiable proof that a user is human and that no duplicate exists, while keeping biometric information on the user’s device. The verifier receives the result of the check rather than the raw personal information used to generate it. Yanez describes the architecture as zero-knowledge by design, with credentials anchored through open cryptographic roots rather than issued exclusively by one central identity provider.

The longer-term aim is to turn human verification into a portable proof. A person could demonstrate something important about themselves—human, unique, authorised—without disclosing everything else.

From identity testing to proof of humanhood

Yanez did not begin with this entire vision fully formed. Subnet 54 was originally known as MIID and focused more narrowly on identity testing and adversarial data generation for fraud prevention.

The MIID repository describes miners generating identity-preserving variations of facial images. These can alter pose, lighting, facial expression, background conditions or the appearance of a screen-replay attack while retaining the relevant characteristics of the original identity. Validators then assess whether those outputs are realistic and useful for improving fraud-detection, KYC and biometric-verification systems.

This addresses a practical weakness in identity security. A verification model cannot learn to resist every future attack from a fixed collection of known examples. Fraudsters continually experiment with new documents, presentation attacks, deepfakes and synthetic identities. Defenders therefore need a stream of increasingly convincing adversarial material against which their systems can be tested.

Yanez’s newer direction expands this work. Instead of concentrating only on helping institutions recognise manipulated identities, the project is moving towards a system that can prove humanhood, uniqueness and authorisation directly.

That is a substantial expansion of the thesis. Generating adversarial identity data and operating a production-grade proof-of-humanhood network are related problems, but they are not identical. The first helps verification systems become harder to deceive. The second must also establish a reliable credential, preserve user privacy, prevent duplicate enrolment and remain trustworthy across institutions and devices.

The connection between the two is nevertheless logical. A proof-of-humanhood system will only be useful if it survives serious attempts to fool it. Yanez’s earlier work gives the subnet a natural role in creating those attacks.

Why build this on Bittensor?

Identity verification is an adversarial field. Once a system becomes widely used, defeating it acquires economic value. Attackers search for weak liveness checks, demographic blind spots, replay vulnerabilities and edge cases that the original developers failed to anticipate.

A conventional identity company responds by maintaining an internal security team and periodically updating its models. Yanez is exploring whether Bittensor can turn that process into a continuous open competition.

Within this model, miners can produce new adversarial identity examples, test liveness systems and search for ways to deceive deepfake or fraud-detection models. Validators evaluate which contributions are genuinely useful. The strongest miners receive rewards, giving independent participants a reason to discover attacks before criminals do.

The resulting feedback loop is well suited to Bittensor’s architecture. As fraud techniques improve, miners have an incentive to produce more difficult challenges. Validators determine which attacks expose meaningful weaknesses. The Yanez models can then be trained or tested against those examples, making the product progressively harder to fool.

The team also describes Bittensor as part of its approach to decentralised proof of uniqueness. Rather than relying entirely on one institution to decide which users are legitimate, the intended network would help enforce the rule that one human can maintain only one valid instance.

This is where the design becomes both more ambitious and more difficult. Producing adversarial examples is relatively easy to express as a competitive task: an output either exposes a weakness or it does not. Establishing global uniqueness without revealing identity is harder. The system must recognise that the same person has already enrolled without creating a central biometric registry that defeats the privacy objective.

Bittensor can provide competition and incentives, but it does not make this cryptographic and governance problem disappear. It gives Yanez a market through which possible solutions and attacks can be discovered. Whether that market can support a robust identity protocol is what the subnet still has to demonstrate.

Where could Yanez be used?

Financial institutions are an obvious market. Banks and fintech companies already spend heavily on detecting identity fraud, manipulated documents, deepfakes, screen replays and biometric spoofing. At the same time, they face growing security and privacy costs from collecting large quantities of personal information.

A Yanez-style system could allow an institution to verify a specific property without storing all the information normally associated with it. A bank might still need to establish a customer’s legal identity during onboarding, but later interactions could rely on narrower proofs: that the same authorised human is present, that a transaction was approved by a real person, or that an AI agent received valid permission to act.

Crypto presents a different version of the problem. DAOs, quadratic funding systems, airdrops and on-chain reputation networks often need Sybil resistance, yet many participants do not want those systems to become conventional KYC databases. Proof of uniqueness could provide a middle layer between complete anonymity and complete identification. The network would know that one distinct human is participating without necessarily knowing who that person is in the civil world.

The more speculative use case involves AI agents.

An agent may eventually pay invoices, manage subscriptions, sign transactions, submit applications or interact with financial services on someone’s behalf. The institution receiving those actions will need more than evidence that the software possesses the correct API key. It will need some way to verify that a real human authorised the agent and to understand the boundaries of that authority.

Yanez describes this through verifiable delegation chains. A human would grant an agent a defined form of permission, while the resulting actions could be traced back cryptographically to that authorisation without requiring the human to expose unnecessary personal information each time.

This could become an important part of the agentic economy. A payment company is unlikely to care that the verification system originates from a Bittensor subnet, and a bank will not adopt it because its tokenomics are interesting. They will care whether it reliably answers a practical question:

Did a real and properly authorised human approve this action?

That is the commercial test Yanez ultimately has to pass.

What still needs to be proven?

The project’s direction is compelling, but several important claims remain to be demonstrated in practice.

Keeping biometric data on the user’s device would reduce one of the largest risks associated with centralised identity systems, although the quality of that protection depends on the actual hardware, software and cryptographic implementation. A privacy-preserving design can minimise disclosure, but users still need confidence that the device is performing the promised checks and that the resulting proof cannot be forged or linked across services in undesirable ways.

Uniqueness is an even more demanding problem. The system must prevent one person from enrolling repeatedly across different devices or after resetting an account, while still avoiding a central database that can reconstruct everyone’s identity. It must also handle false positives, false negatives, changing appearances, inaccessible hardware and users who cannot complete conventional biometric checks.

The decentralised component requires scrutiny as well. It matters which tasks miners actually perform, how validators measure quality, which actors can update the proof system and whether a small number of participants can influence enrolment or verification. Calling an identity protocol decentralised does not automatically remove its trust assumptions; it may simply distribute them across hardware providers, validators, smart contracts and cryptographic infrastructure.

Commercial adoption remains the final filter. Recent reporting around Yanez has linked revenue from its newer products to purchases of the Subnet 54 alpha token. That could strengthen the relationship between product demand and the subnet economy, but buybacks are secondary to the more important question of whether institutions are willing to pay for the verification service itself. Emissions can help a network develop. They cannot permanently substitute for customers.

Why Yanez is interesting

Many Bittensor subnets compete around familiar forms of AI production: inference, training, image generation, data or model evaluation. Yanez applies the incentive structure to a less visible but potentially foundational problem.

As synthetic media and autonomous agents improve, it will become increasingly difficult to infer human presence from online behaviour. Writing style, facial appearance, voice and account history will all become weaker signals. The internet will need more reliable ways to establish when a real person exists behind an action, and when that person has delegated authority to a machine.

The uncomfortable response would be to expand centralised identity systems until every meaningful online interaction requires a platform, government or corporation to inspect and store more personal information. Yanez is exploring another route: prove only what the interaction requires, keep the underlying data with the user and use an adversarial market to make the verification system progressively harder to deceive.

Whether Yanez can achieve that vision remains uncertain. Proof of uniqueness is technically difficult, institutional identity markets are conservative, and privacy claims must survive more than a well-designed product page. But the direction is structurally important.

The internet has learned how to recognise devices, credentials and accounts. In an environment filled with bots, deepfakes and autonomous agents, it will increasingly need to recognise legitimate human authority as well.

Yanez is trying to provide that proof without first turning the human into a permanent entry in someone else’s database.

More information: https://www.yanez.ai/