AI Is Helping Discover Tech Vulnerabilities—And Zcash Is Just the Latest Example

The latest generation of frontier AI models are no longer just chatting with users, generating images, or writing code. Researchers are increasingly using systems such as Anthropic’s Claude Mythos and Claude Opus 4.8 and OpenAI’s GPT-5.5 to identify software vulnerabilities, raising concerns about what happens when those capabilities become widely available.

Crypto investors got a wake-up call about the rising threat from powerful AI this week when Zcash developers disclosed that Claude Opus 4.8 helped discover a critical vulnerability that could’ve enabled an attacker to mint unlimited ZEC. Due to the network’s design, there’s no current way to know for sure whether counterfeit ZEC was, in fact, minted—and that uncertainty led to the price of ZEC crashing late this week.

Experts warn that many more vulnerabilities could be found in the coming weeks and months as AI software gets more capable—and those tools become more accessible. Here’s a look at the growing threat, and how it’s already impacted the crypto world.

Early AI models were professionally used as coding assistants, helping developers write, explain, and debug software. As the technology improved, researchers began using the same systems for code review, software auditing, and vulnerability research.

The transition from coding assistant to security tool coincided with a broader shift in how AI was being used inside software development. After the launch of Claude Code in 2025, Anthropic reported a sharp increase in AI-generated code across its engineering teams, reflecting a move from models that suggested code to systems capable of writing and running it.

Security professionals say the implications extend beyond helping developers write code.

“AI is far better at reviewing code than most people and finding potential vulnerabilities in it,” Danny Jenkins, CEO and co-founder of ThreatLocker, told Decrypt. Jenkins said current AI systems are already accelerating vulnerability discovery, while newer models such as Mythos could significantly expand those capabilities, calling it an imminent “big problem.”

“It will be only a matter of time until someone bad gets access to it,” he said.

According to Jenkins, AI is also lowering the barriers to entry for vulnerability research, allowing more people to analyze code, identify weaknesses, and develop exploits. As access to increasingly capable systems expands, he expects the pace of vulnerability discovery to increase.

“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.”

As AI systems became more capable, companies began applying them to cybersecurity. On Tuesday, Anthropic expanded access to Project Glasswing, giving 150 companies and institutions access to Claude Mythos to help identify and remediate software vulnerabilities before the model is released more broadly.

In April, Mozilla later disclosed that Anthropic’s models helped identify hundreds of vulnerabilities that it fixed in the Firefox web browser, while researchers at Calif used Mythos Preview during work that produced one of the first public exploits targeting Apple’s M5 chips.

Stanislav Fort, a former researcher at Google DeepMind and Anthropic and now founder and chief scientist of security firm Aisle, said concerns about AI-powered vulnerability discovery are valid, but often misunderstood.

“The naive response is to try to gatekeep access to powerful models. I think this is essentially security by obscurity, and security by obscurity is one of the worst ideas in the field,” Fort told Decrypt. “The capability for zero-day discovery is already widely distributed across models that no one can restrict. Trying to bottle it up at the frontier doesn’t eliminate the risk; it just delays it while also slowing down the defenders who need these tools most.”

Fort said the greater risk is that defenders, particularly open-source maintainers, may lack access to the same advanced AI tools available to attackers.

“That imbalance is the real danger,” he said. “The answer isn’t restriction; it’s democratization of the defensive stack.”

Anthropic is not alone in pushing AI models aimed at cybersecurity. In May, Microsoft introduced MDASH, an agentic vulnerability discovery system that the company said helped identify previously unknown Windows vulnerabilities.

The risk to crypto

Crypto and DeFi are starting to feel the impact of AI-powered bug hunting. Blockchain projects have always been attractive targets because there is a lot of money at stake and much of the code is publicly available. Jenkins said as AI gets better at finding software flaws, open-source crypto projects could become easier targets for both security researchers looking for bugs and attackers looking to exploit them.

In one of the clearest examples of how advanced AI models can help researchers uncover vulnerabilities that had survived years of human review, independent security researcher Taylor Hornby disclosed the critical vulnerability in Zcash’s Orchard privacy pool that he discovered with the assistance of Claude Opus 4.8.

The flaw could have allowed an attacker to create unlimited counterfeit ZEC, and had gone undetected for years before being patched. Whether the exploit was actually used currently remains unknown.

“The vulnerability was present from Orchard’s activation in May 2022 until the emergency fix was deployed on June 1, 2026,” Shielded Labs, the organization behind Zcash development, wrote in a disclosure post. “Due to the privacy properties of Orchard and the nature of the bug, there is no definitive way to determine, using only cryptography, whether such exploitation occurred.”

The attack comes as DeFi protocols are already facing one of their worst years for exploits. More than $840 million was stolen from DeFi projects in the first five months of 2026, including more than $600 million in April alone across attacks on projects including KelpDAO, and Drift Protocol.

The rise of so-called ‘vibe hacking,’ where attackers use AI coding agents to automate reconnaissance, credential theft, malware development, and other tasks, has raised concerns that AI is lowering the barriers to carrying out sophisticated cyberattacks

According to Natalie Newson, senior blockchain investigator at Web3 security platform CertiK, while April was unusually severe for crypto exploits, the broader trend remains more stable and below the peak number of incidents seen in past years.

“April 2026 was a bad month for crypto exploits; there were only three days without an exploit in which at least $10,000 was taken,” she said. “However, when we take a look at the wider picture, the number of incidents (excluding phishing) has arguably been fairly consistent and still lower than a peak in 2023.”

While AI is making DeFi exploits easier to carry out, according to Blockaid CTO Raz Niv, the bigger risk is not AI replacing hackers but amplifying them, allowing attackers to focus on more sophisticated techniques while AI handles routine tasks.

“The good news is defenders can use the same tools,” he said. “AI-assisted monitoring and simulation is becoming essential for security teams trying to keep pace.”