Vitalik Buterin: AI Will Not End Crypto Security
Ethereum co-founder Vitalik Buterin published a rebuttal to the claim that "AI hackers will inevitably defeat cybersecurity," stating that cybersecurity naturally favors the defensive side after serious defensive measures are taken, and that AI will not cause a collapse in crypto security.
He emphasized formal verification as a core approach: using mathematics to prove that software meets predefined security properties, rather than relying on manual code reviews to gamble on the absence of flaws. His analogy was that if AI could prove the Navier-Stokes equations and Fermat's Last Theorem, it could also prove that "this program is secure," even if the program itself is complex. He acknowledged that the definition of "security" can easily be narrowed, potentially overlooking protocols, servers, databases, network layers, caches, and supply chain dependencies, which could lead to a false sense of security.
He tied his own position to this judgment: about 90% of his net worth remains in crypto assets, asserting that continuing to hold crypto is a bet that defenses can keep pace with stronger automated attacks. Earlier, he dismissed predictions that AI would reduce Bitcoin's value by 50% within two years, believing the probability of AI fundamentally breaking hash and proof-of-work is extremely low. He noted that layers like clients and mining pools, which can be patched without societal consensus, are the transitional challenges, and the same logic applies to Ethereum.
Ongoing projects are being developed: arklib is advancing the full verification implementation of Stark, evm-asm is rewriting the entire EVM in assembly with correctness proofs, and both Vyper and Verity smart contract languages are working on formal verification. The Ethereum Foundation's protocol security team has used AI agents to scan protocol code and independently reproduce vulnerabilities, confirming a remotely triggerable crash in rust libp2p gossipsub, numbered cve-2026-34219, triggered by a specially crafted prune message causing a backoff computation overflow, which has been fixed in version 0.49.4.
On the research side, verification is being made a cross-cutting tool rather than a point audit. The Foundation's September protocol priorities have laid out formal verification for privacy, state, zkEVM, and post-quantum tracks extending to 2029; the better.codes initiative launched in August allows researchers to write cryptographic soundness issues into Lean, with the Lean kernel checking proofs against fixed theorems; funding reports also include etheorem comparing clients with specifications, speca with leanagent for protocol compliance, and RISC-V zkvm related verifications. He emphasized that without a path for scalability and privacy chains, there is no future; verification must cover the entire stack and cannot just focus on "key modules."
Both offense and defense are accelerating. Anthropic disclosed that its cutting-edge models have accelerated vulnerability research, exploit development, and multi-target campaigns, having disclosed 2,300 AI-discovered vulnerabilities across 392 open-source projects by August 26, with 421 already patched upstream; its threat intelligence has also recorded AI exploitation pipelines targeting security devices and government entities. The same model is used on the defense side for pre-disclosure checks. He cited an example from 2026: a post-quantum cryptography library claiming to be "verified" was still breached, usually due to incomplete verification coverage or missing key properties in specifications. Formal verification, testing, and type systems must be used in conjunction; a single tool cannot guarantee full coverage.
In market mechanisms, this is an event-driven narrative trade rather than an immediate capital shift: continuing to hold coins equates to betting that "defense costs decrease faster than attack costs decrease," benefiting Ethereum L1 research, zkEVM, Stark toolchains, and verification languages; unverified contracts, supply chain dependencies, and user intent layers (phishing, misauthorization, agent rogue signing) are under pressure. The offensive side is purchasing cheaper vulnerability pipelines, while the defensive side is acquiring machine-checkable specifications and proofs; capital will not immediately shift due to a single post, but the security premium will transition from "audit report page count" to "whether the entire stack is verifiable." No significant ETH flows directly corresponding to this statement have been observed on-chain, and price reactions remain primarily at the narrative level.
Source: Public Information
ABAB AI Insight
Since writing the Ethereum yellow paper in 2013-2014, Vitalik has treated "irreversible funds + open attack surface" as a design constraint rather than a post-factum public relations issue. The hard fork after the DAO hack, multiple client consensus incidents, and his repeated push for minimal trust upgrades all follow the same path: first acknowledging that software can fail, then narrowing the scope of failure from "societal voting" to "verifiable specifications and interchangeable clients." In May 2026, he referred to AI-assisted formal verification as the "ultimate form" of software development, specifically mentioning quantum-resistant signatures, Stark, consensus algorithms, and zk-EVM; this statement pushes the same technical route from research blogs to balance sheets.
The capital path is not about opening new funds but continuing to direct the Foundation's years of budget, protocol security group, and Lean proof stack in the same direction. Money flows into three types of assets: cryptographic primitives that can be machine-checked (Stark, post-quantum), provable execution layers (zkEVM, evm-asm), and tools that expose "specification errors" (etheorem, better.codes). The motivation is strong: Ethereum needs to add both privacy and scalability, with the attack surface increasing exponentially with code volume, and the marginal cost of manual audits has already hit a ceiling; if AI reduces the cost of writing code, but does not also reduce the cost of proofs, L1 will become a "factory for generating vulnerabilities." Having 90% of net worth in crypto serves as a guarantee for this R&D path rather than for a specific token's market.
Analogies must be grounded in specific companies. Intel and management have used formal methods to tackle microarchitecture correctness, seL4 has mathematically proven microkernel correctness, and AWS has used formal tools to check cryptographic libraries and authorization logic; the crypto side is closer to "irreversible funds of seL4." The current industry phase is not about expanding user numbers but transitioning from "audit + bounty" to "specification + proof" control: whoever can reduce the cost of full-stack verification will take the pricing power of the next generation settlement layer. Bitcoin follows a conservative path of making fewer changes to consensus and more to clients; Ethereum is embedding verification tools into its long-term roadmap. Both sides are betting on the same thing: cryptographic primitives are more resistant to AI than application layers, and the real vulnerabilities that will be amplified by AI are specification flaws and human-machine interfaces.
Structural judgments belong to the overlap of technical substitution and pricing power transfer. What is being substituted is not crypto itself, but the production function of "humans reading code to find bugs": the offensive side continuously scans the exploitation chain with models, while the defensive side must shift to "first write properties, then let machines falsify them." Pricing power is shifting from audit firms and post-factum compensation to protocols and toolchains that can produce verifiable proofs. The mechanism layer reason is simple: vulnerability discovery is a one-time event, while proofs, once established, can be reused; as the cost of generating attacks approaches zero, only by changing "security" from an empirical judgment to a verifiable theorem can open networks continue to host irreversible funds.
ABAB News · Law of Cognition
- Attack once, proof reusable once.
- The cheaper code generation, the more expensive unproven systems.
- Position is opinion, specification is defense.