Zcash founder Zooko Wilcox said an AI-assisted security audit found no new serious vulnerabilities in the privacy-focused cryptocurrency’s protocol after developers fixed a previously discovered Orchard bug.
The audit was commissioned by Shielded Labs, a Swiss non-profit foundation that supports Zcash development. According to Wilcox, Anthropic’s Claude Mythos model reviewed the Zcash protocol and did not identify any additional serious flaws.
The result came after a tense moment for the Zcash ecosystem. On June 3, developers temporarily paused transactions in the Orchard shielded pool after discovering a vulnerability in its design. Functionality was restored later the same day through an emergency upgrade.
The bug had reportedly existed for about four years and was discovered by security researcher Taylor Hornby with assistance from Anthropic’s Claude Opus 4.8 model. The Zcash Foundation said it found no evidence that the flaw had been exploited, no signs of unauthorized ZEC creation, and no indication that users’ private data had been exposed.
The Zcash case highlights the growing role of AI in crypto security. Advanced models are increasingly being used to search for flaws in complex protocols, smart contracts, and cryptographic systems that are difficult for humans to review manually.
But the same technology is also creating new risks.
Anthropic recently released Claude Fable 5, a public version connected to its Mythos-class security research models. The launch followed claims that Mythos had identified thousands of high- and critical-severity vulnerabilities in systemically important software.
That raised an uncomfortable question for the cybersecurity industry: if AI can help defenders find vulnerabilities faster, it can also help attackers do the same.
Anthropic said Fable 5 had been adapted for general use and included safety mechanisms. Some requests involving high-risk areas, including cybersecurity, were redirected to Claude Opus 4.8. But shortly after release, access to Fable 5 and Mythos 5 was suspended following pressure from US export-control regulators, who cited national security concerns.
The move showed that governments are starting to treat advanced AI security models not just as software products, but as potentially sensitive technology.
The clean Zcash audit is a positive signal for the project, but it does not fully resolve the deeper issue raised by the Orchard bug.
Zero-knowledge proof systems are mathematically complex, and their security depends on subtle constraints that can be extremely difficult to verify through traditional review. In that environment, AI may have a structural advantage: it can search through large codebases, reason across technical dependencies, and test hypotheses at a scale that human auditors cannot easily match.
That is why the Zcash incident matters beyond one protocol. It suggests that AI-assisted auditing may become a standard part of crypto security, especially for privacy systems and other highly technical infrastructure.
The Hard Question Is What AI Still Cannot Prove
The biggest unresolved issue is not only whether AI can find bugs. It is whether AI can prove that no serious bugs remain.
Because Orchard is private by design, it is difficult to retrospectively audit the full ZEC supply over the period when the bug existed. Shielded Labs is therefore working on a proposal for a “turnstile” mechanism that would help verify the integrity of the coin supply.
This is where the limits of AI auditing become clear. AI can help discover vulnerabilities before attackers exploit them, but it cannot automatically remove structural uncertainty from systems where privacy prevents full historical visibility.
For Zcash, the latest audit is reassuring. For the wider crypto industry, it is a warning. AI is quickly becoming one of the most important tools in security — but it is also lowering the barrier for attackers and forcing regulators to pay closer attention to who can access the most powerful models.
The next phase of crypto security may not be defined only by better audits. It may be defined by who gets to use the AI models capable of finding the bugs first.
