Frontier AI Models Can Find Crypto’s Biggest Bugs. Experts Warn the Industry Isn’t Ready

A security researcher using Anthropic’s Claude Opus 4.8 uncovered a critical flaw in Zcash’s Orchard privacy pool in a matter of days, exposing a vulnerability that had survived four years of review by leading zero-knowledge cryptographers.

The disclosure sent ZEC tumbling roughly 38% on Thursday and raised a broader concern for the crypto industry around frontier AI models becoming increasingly proficient in finding vulnerabilities than most humans.

“The significance isn’t really that AI can find bugs,” Ben Goertzel, founder and CEO of SingularityNET, told Decrypt. “It’s that the kind of bug it can now find has changed.”

Rather than simply flagging obvious coding mistakes, frontier models are increasingly capable of reasoning about whether software behaves the way its designers intended, he said.

In May, Taylor Hornby, a security researcher hired by Shielded Labs, discovered a critical flaw in Zcash’s Orchard circuit with assistance from Anthropic’s Claude Opus 4.8. Hidden in two lines of code, the bug stemmed from a check that appeared to validate transaction inputs but wasn’t actually enforcing the intended rules, potentially allowing an attacker to create counterfeit ZEC inside the shielded pool without detection. Hornby built a working exploit to verify the vulnerability before reporting it to developers. An emergency fix was deployed on June 1.

Adding to the panic that hit Zcash and the broader crypto market on Thursday and Friday is the fact that the flaw had been left undiscovered for over four years.

For Goertzel, the discovery is significant not only because AI found a vulnerability, but also because it points to a new model for security research.

“I think it’s an early marker of a shift that’s going to be hard to overstate,” he said. “The model of security research as a handful of revered human specialists doing slow, artisanal, deeply-expert audits doesn’t go away, but it stops being the whole game.”

Goertzel said the Orchard flaw belongs to a class of subtle logic bugs that frontier AI models are increasingly capable of finding, including smart-contract errors, access-control failures, and situations where software behaves differently than its designers intended. As those capabilities improve, he added that security research is shifting toward a model in which human specialists oversee continuous AI-driven review that can analyze codebases far more extensively than traditional audits.

The Zcash response itself may offer a preview of that future, Goertzel said.

“Shielded Labs bringing on a researcher specifically to hunt protocol-level flaws with a frontier model before a malicious actor could is, I suspect, the template, not the exception,” Goertzel said. “Proactive, AI-augmented, adversarial-by-design review becomes table stakes, and the protocols that don’t adopt it will increasingly be the ones learning about their vulnerabilities from the attacker rather than from a friendly.”

According to Sean Ren, CEO of Sahara AI and a computer science professor at the University of Southern California, advances in AI are also reshaping the balance between attackers and defenders as frontier models can rapidly test attack strategies, learn from the results, and uncover weaknesses.

“In order to build up better defense, we have to use these frontier AI models as the potential attackers to stress test these systems,” Ren told Decrypt.

Ren said blockchain networks are especially exposed because their open-source code can be analyzed directly by frontier AI models, which can rapidly test attack strategies and identify vulnerabilities faster than traditional security reviews.

“If you think about frontier model labs like OpenAI, Anthropic, and Google DeepMind, they have earlier access to the strongest unpublished models and can conduct a lot of experiments on public network systems like blockchains, so they do have the power at hand,” he said. “If someone with malicious intent had access to those capabilities, they could conduct attacks and create vulnerabilities.”

That window may close faster than many expect, and according to Danny Jenkins, CEO and co-founder of cybersecurity firm ThreatLocker, AI-assisted vulnerability discovery is improving faster than many organizations can secure the software they already rely on.

“We have this huge gap that’s going to take years and years to get through,” Jenkins told Decrypt. “All of this software is going to have all of these vulnerabilities, we’re not going to have fixes or updates for it for a long time, and people are going to be able to find those vulnerabilities very quickly.”

Jenkins said AI is not fundamentally changing vulnerability research so much as dramatically accelerating it. Tasks that once required security researchers to review code and reverse engineer software manually can now be performed in seconds by modern models.

“Pre-AI, cybersecurity threats and exploits were increasing every year,” he said. “Post-AI, it’s become even faster, and I think it’s become faster for two reasons. One is that you can now use AI to help find vulnerabilities and exploits, and the number of people who have the ability to do this has massively grown. You don’t have to be a script kiddie now.”

Despite those risks, Goertzel argued that crypto may also be better positioned than other industries to adapt because its code is open, and its communities are highly security-focused.

“Crypto is standing closest to the door, but it’s also the part of the room that can see the door coming,” he said.