AI Discovers Long-Dormant Zcash Bug for 4 Years

Artificial intelligence (AI) has uncovered vulnerabilities and attack vectors in previously analyzed open-source cryptocurrency code that human researchers missed over years. In the case of Zcash (ZEC), a flaw enabling unlimited counterfeit coin production was discovered, while Coldcard experienced Bitcoin theft linked to an outdated firmware error.

Realized Labs announced on May 29 that AI-assisted audits identified a vulnerability in Zcash's Orchard circuit. The issue, termed a 'soundness failure,' allowed attackers to create undetectable counterfeit ZEC by generating data resembling valid proofs. The audit utilized Anthropic's Claude Opus 4.8 and custom audit agents. Attack code was verified in a local test environment, prompting the development team to complete emergency fixes by June 2.

The Orchard pool is Zcash's privacy transaction zone. Enhancements like Orchard pool restrictions and Ironwood pool implementation aim to limit funds that could leak externally if a zero-knowledge proof error occurs. This case does not indicate AI broke new cryptographic systems but rather quickly identified logic errors in open code, validating attack feasibility.

Coldcard's incident involved a 2021 firmware random number generation flaw later exploited. CoinKite reported confirmed losses of 1,778.84 BTC, over 8,600 addresses, and $112.7 million (approximately KRW 156.3 billion). The vulnerable firmware existed since 2021, with Bitcoin being stolen from July 2026 onwards across multiple instances. It remains unclear if AI was used in the attack.

Another case involved Bankr on May 4, where attackers posted Morse code on X (formerly Twitter), which Grok interpreted as commands, triggering a bot to send approximately $3 billion DRB. Agent Wormhole estimated damages at around $150,000–$200,000 at the time. The core issue was permission design rather than cryptography: AI translated external input into transaction commands without sufficient verification or spending limits.

Smart contracts also pose risks when source code is unverified. Chainalysis reported that over $36.7 million (approximately KRW 50.9 billion) was stolen from unverified smart contracts in the last six months. Attackers must decompile bytecode to find vulnerabilities, but AI tools may lower these reverse engineering costs.

Blockchain 'dead drops' for hiding malicious command-and-control data have increased by 440%, rising from an average of 2.06 daily records before high-performance open-source Chinese AI models emerged to 11.1. This does not indicate AI executed individual attacks but suggests lower barriers for building attack infrastructure.

AI has also been used in fraud promotions, with TRM Labs noting nine YouTube videos claiming 'Claude-created arbitrage bots' led to 224 individuals sending 274.60 ETH ($517,205) to 234 malicious smart contracts between February and August 2026. No evidence suggests AI decrypted Bitcoin (BTC) cryptography in these cases.

Security audits must separately verify vulnerability discovery, actual exploitation, loss magnitude, and whether AI was used by attackers. For Zcash, counterfeit creation was confirmed in a test environment but not on mainnet; Coldcard losses were reported without confirming AI involvement. All cases highlight that security checks should not assume risks disappear after initial audits.