Technology•

Aztec Network CEO Zac Williamson Says Newer AI Models Could Still Expose Bugs in Tierra

Zac Williamson says Aztec has formally verified only its most critical code and forces users to migrate funds at each major upgrade so hidden double-spends cannot carry over.

Written By :MINRK
Aztec Network CEO Zac Williamson Says Newer AI Models Could Still Expose Bugs in Tierra

Photo by Shahadat Rahman on Unsplash

Mentioned in This Article

ETH
ETH
Ethereum
$1,807.95
0.61%
ZEC
ZEC
Zcash
$890.19
0.89%

1. Aztec Network Chief Concedes AI Could Still Find Flaws in Tierra

Zac Williamson, co-founder and chief executive of the Aztec Foundation, said in an interview on November 9 that he cannot rule out more advanced AI models uncovering bugs in Aztec Tierra, the Alpha V6 release of the Aztec Network privacy layer on Ethereum that went live on November 3, 2026.

The admission is notable because it comes from the head of a project that has leaned heavily on artificial intelligence to harden its own code. Williamson did not claim that Tierra is free of defects. He said instead that the team cannot exclude the possibility that stronger models will surface problems the current tools missed, a candid position for a network whose central promise is keeping user activity private and user funds safe.

His remarks sketch a security approach built on three layers: continuous AI-driven auditing, mathematical proofs for the most sensitive parts of the code, and an upgrade process designed so that any undetected flaw in one version cannot be carried into the next. Each layer addresses a different failure mode, and together they reflect how privacy networks are rethinking risk as automated bug hunting becomes cheap and widely available.

Aztec occupies an unusual position in the Ethereum ecosystem. Most layer-2 networks focus on throughput and lower fees, publishing transaction data openly so anyone can audit balances. Aztec aims to let users transact without exposing amounts or counterparties, which makes the integrity of its cryptography far more consequential. In a transparent system, an accounting error tends to be visible. In a private one, it can hide.

2. What Did AI Audits Find in Earlier Versions?

AI audits of the previous Aztec release turned up "some criticals" in the network's cryptographic libraries, Williamson said. Those findings were the most serious category of vulnerability, the kind that can undermine the guarantees a protocol is meant to provide.

The most consequential example became public in August. Aztec disclosed then that its internal AI-assisted auditing had identified a critical flaw in the proving system used by Version 5. The weakness could have allowed a forged proof to be accepted as valid. In a zero-knowledge system, proofs are the mechanism that convinces the network a transaction follows the rules without revealing its contents, so a proof that can be faked strikes at the foundation of the design.

Aztec said at the time that it could not determine whether anyone had exploited the flaw. That uncertainty is a recurring feature of privacy technology: the same properties that shield user data also make it difficult to reconstruct, after the fact, whether a vulnerability was abused.

On October 15, a post co-written by Williamson stated that the AI audit harness had caught critical bugs that human auditors had missed, and that V6 contains the patches. Aztec's launch announcement for Tierra confirmed that the release fixes the V5 proving flaw. The same announcement introduced an AI testing harness and described formal verification of core cryptographic components.

3. How Aztec Runs Its AI Testing Harness

Aztec's AI audits run around the clock on infrastructure the team built itself, Williamson said. Rather than commissioning a review at fixed points before a release, the project keeps automated systems probing its code continuously.

Williamson framed that choice in terms of the threat landscape. The cost of mounting attacks, he argued, is falling toward zero, and adversaries could soon deploy millions of AI bots to search for weaknesses. If attackers can scan code at that scale for almost nothing, a defender relying on periodic manual reviews is at a structural disadvantage. Running equivalent tooling internally, and running it constantly, is an attempt to find the flaws before someone hostile does.

He also addressed a practical obstacle that security researchers sometimes encounter with commercial AI models: refusals, where a model declines a request it judges potentially harmful. Williamson said the team rarely runs into that problem, because its work centers on cryptography and general software bugs rather than on requests that look like offensive hacking. As an additional safeguard, Aztec also runs uncensored open-source models to cross-check results, giving it a second set of eyes that is not shaped by the same guardrails.

The approach reflects a broader shift in how software security is practiced. Traditional audits are point-in-time snapshots performed by a limited number of specialists. Automated harnesses can cover more code paths, repeat tests after every change and operate without fatigue. Williamson's comments suggest Aztec now treats AI auditing as a permanent part of development rather than a one-off exercise, while acknowledging that tomorrow's models may see what today's do not.

4. What Is Formal Verification and How Much of Aztec Has It?

Formal verification is the practice of mathematically proving that code behaves exactly as specified, and Williamson said only the most security-critical parts of the Aztec protocol have been verified this way so far. The full protocol has not yet been formally verified.

Unlike testing, which checks behavior against a finite set of inputs, a formal proof establishes that a piece of code meets its specification for every possible input. It is demanding work, usually requiring specialized tools and expertise, which is why projects typically apply it selectively to the components where an error would be most damaging.

Aztec's sequencing illustrates that logic. After its AI tools surfaced critical bugs in particular subsystems, the team moved to formally verify those subsystems. Williamson said this implies the bug the AI found was the last one in that part of the code: once a component is proven correct against its specification, no further defects of that kind can remain there.

The limitation is the parts of the system that have not yet received that treatment. That gap is precisely why Williamson stopped short of claiming Tierra is bug-free. AI auditing can find many problems, but it cannot prove their absence. Formal verification can, but only where it has been applied.

Williamson said he sees formal verification as the long-term solution. In his view, the end state for a privacy network is code whose correctness is established by proof rather than by the absence of discovered bugs.

5. Why Hidden Double-Spends Worry Privacy Networks

Bugs in privacy networks can enable double-spends that no one observes, Williamson said, which is what makes them more dangerous than many flaws in transparent blockchains.

A double-spend occurs when the same funds are used more than once, effectively creating value from nothing. On a public ledger, balances and transfers are visible, so inflation of this kind can usually be spotted by anyone tracking supply. In a shielded system, where amounts and parties are hidden by design, a successful double-spend could go unnoticed, quietly diluting the holdings of every honest user.

The topic came up in connection with a Zcash vulnerability that became public in June. Zcash is one of the longest-running privacy-focused cryptocurrencies, and its developers have long had to think about how to detect counterfeiting in pools where transactions are encrypted. The project's advisory on that bug said its turnstile accounting capped how much value could leave the affected pool.

A turnstile, in this context, is a mechanism that tracks the total value entering and exiting a shielded pool. Because those flows cross the boundary between the private and public parts of the system, they can be counted even when individual transactions inside the pool cannot be seen. If more value tries to leave than ever went in, the discrepancy is exposed, and the damage any hidden inflation can do is bounded.

6. How Aztec Uses Hard Forks to Contain Hidden Bugs

Aztec borrows the turnstile idea from Zcash and applies it through its upgrade process, Williamson said, praising the way Zcash manages risk. For now, every major Aztec upgrade is carried out as a full hard fork rather than an in-place update.

In practice, users must withdraw their assets from the old version of the network and redeposit them into the new one. That migration acts as a checkpoint. If a double-spend had been hidden inside the previous version, it would reveal itself when users withdraw Ethereum-based tokens back to the base layer, because the old system would be unable to honor more withdrawals than it actually holds.

Crucially, a hidden flaw cannot follow users into the new release. Each version starts with only the funds that were legitimately withdrawn and redeposited, so any counterfeit balance created through an undiscovered bug stays behind. The cost of the design is convenience: users must actively move their funds with each major upgrade. The benefit is a hard boundary on how far an undetected error can spread.

The approach fits the broader philosophy Williamson described. Aztec accepts that it cannot yet prove every line of its code correct, so it structures the system to limit the consequences of the mistakes it may not have found.

7. Independent Rating and the Absence of a Security Council

ETHBEAT, the Ethereum layer-2 tracker formerly known as L2BEAT, lists Aztec V6 as a fresh start that carries over no state from V5. The tracker rates V6 at Stage 2, its highest tier.

Stage ratings are used across the Ethereum ecosystem to describe how far a layer-2 network has moved away from reliance on trusted operators. A Stage 2 designation signals the most mature level of decentralization in that framework, meaning users depend least on any small group to keep the system honest.

Williamson pointed to one governance difference in particular: Aztec has no security council. Many other networks keep such a council, typically a small group of signers with the power to intervene in an emergency, for example by pausing contracts or pushing urgent fixes. Those councils offer a safety net but also introduce a point of trust and potential control.

Without one, Aztec cannot rely on emergency intervention to rescue users from a catastrophic bug. That makes the containment design described above, along with continuous auditing and formal proofs, all the more central to its security model. The fresh-start status that ETHBEAT records is, in effect, the hard-fork migration policy viewed from the outside.

8. What Comes Next for the Aztec Network

Tierra is the latest stage of a network that is gradually building out its user-facing activity. In September, Aztec Labs relaunched its zk.money payments wallet on the network, giving users a practical way to make private transfers on the new infrastructure.

The near-term roadmap, as Williamson described it, rests on continuing what the team is already doing: keeping its AI harness running around the clock, extending formal verification beyond the most security-critical components, and using hard-fork upgrades as a firewall between versions until the codebase is proven to a higher standard.

His central message was one of measured caution. AI has already found serious bugs that human reviewers missed, and Aztec has fixed them. But the same technology is improving quickly, and Williamson acknowledged that newer models could find issues that today's cannot. For a network whose value depends on keeping private balances both confidential and sound, he argued that mathematical proof, rather than ever more testing, is ultimately the way to close that gap.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY