Johns Hopkins cryptographer Matthew Green has joined the growing chorus of warnings about AI-driven cryptographic threats, warning that AI could discover mathematical weaknesses that decades of human research have missed. Green clarified that he knows of no imminent cryptographic breakthrough, but expressed concern about potential future discoveries. The warning follows OpenAI's October 6 announcement that an unreleased AI model had produced 722 mathematical manuscripts, covering 372 groups of results, after attempting roughly 4,000 unsolved problems. Green noted that most of the open problems solved by OpenAI were actually rather close to a solution, conceptually speaking, and that mathematicians working in the field would have come to solutions soon with new tools in their hands. For now, there is no publicly demonstrated AI attack capable of breaking Bitcoin or Ethereum's core cryptography, but the concern centers on whether the industry can upgrade security quickly enough if AI eventually discovers one.
Vitalik Buterin has warned that AI-driven mathematical breakthroughs could weaken lattice-based cryptography within two years and potentially break the Elliptic Curve Digital Signature Algorithm (ECDSA) behind Bitcoin and ETH wallets sooner than expected. Buterin pointed to the factoring problem, where the number field sieve shortcut forced encryption keys to grow over decades, arguing that if AI delivers 50 years of math in two years, lattices could face similar vulnerabilities. Drake cited 722 mathematical results published by OpenAI as evidence of rapid AI-driven progress, defining a break as recovering a private key in about a week with available hardware such as a large GPU cluster. Modern cryptography is generally designed around extremely high security levels, with Green suggesting AI-driven discoveries could reveal that some systems believed to offer 128-bit security actually provide only 96 or 108 bits. A reduction of 20 bits would make an attack roughly one million times cheaper, though Green questioned how security experts would determine the right key sizes. Even post-quantum cryptography could face questions, as the US National Institute of Standards and Technology approved new post-quantum standards in 2024, including ML-KEM and ML-DSA, which rely on lattice-based mathematics.
According to a new CertiK Intel3D report, agentic AI systems are moving beyond flagging suspicious activity to performing autonomous security operations with limited human intervention. As reported by CertiK, these systems can reason through multiple steps, call external tools and APIs, collect evidence, take actions in live environments and assess results before deciding what to do next. The emerging model represents an AI security workforce where autonomous systems operate within defined roles while human experts move toward supervision, quality control and accountability. Across Web2 and Web3, autonomous systems now triage threats, trace illicit funds, and draft regulatory filings, making governance and human accountability the new prerequisites for adoption. The industry has long viewed quantum computers as the real threat, but this is still years away, forcing it to deal with AI advancement now.
Justin Drake's call to prepare for an AI break of wallet cryptography drew a mixed response as bitcoin slipped below $83,000. Drake used a post on X viewed almost 4 million times to urge the industry to begin planning for what he called 'bunker mode'. Not everyone agreed on the urgency - Ethereum co-founder Vitalik Buterin said the risk from AI-accelerated math should be taken seriously, but advised holders against scrambling to move funds. Samson Mow, CEO of bitcoin technology firm Jan3, told followers there was no need to panic 'because an Ethereum researcher is saying silly things'. The debate highlights the tension between preparing for potential security risks and maintaining normal market operations. Haseeb Qureshi, managing partner at VC Dragonfly, noted that 'doomerism has now hit cryptography', stating that the risk is not quantum but conventional mathematics overturning unproven cryptographic hardness assumptions. Dankrad Feist, another Ethereum researcher, took an extremely pessimistic view, saying if AI cracks blockchain, everything will be worthless, arguing that if elliptic curves are broken, having crypto assets in 'bunker mode' is not going to help as they are still worth zero. Blockstream CEO Adam Back called the warning 'dubious,' with Blockstream researchers having developed SHRINCS, a hash-based signature system using SHA-256.
The automation push comes as crypto losses reached $768.4 million in September across 97 incidents under CertiK's methodology, with total losses during 2026 at roughly $2.68 billion by the end of September. As reported by CertiK, flash loan exploits can drain protocols within seconds while stolen assets move through multiple addresses, bridges and mixing services within hours. The company's H1 2026 security findings put crypto losses at $1.32 billion during the first six months of the year, down 46.8% from a year earlier, while wallet compromises became the largest attack method during the second quarter. Bitcoin also relies on elliptic curve signatures, with Drake warning that ECDSA could break within months in the worst case scenario. Drake pointed to Project11's 'risq list,' a tracker of Bitcoin public keys already exposed onchain, noting that wallets holding under 50 BTC get partial cover from approximately 20,000 early exposed addresses holding 50 BTC each. Project Eleven's Bitcoin Risq List tracked more than 8.1 million BTC in addresses it classified as quantum-vulnerable in its September 14 update, with Drake specifically urging major exchanges including Binance, Bitbank, Robinhood, Bitfinex and Tether to examine their cold-storage setups. Some investors are already taking precautions - Charles Edwards, founder of Capriole Investments, said he has stopped using DeFi and reduced his number of wallets and accounts, stating that if AI compromises mainstream encryption in any way, there will be chaos.
