Vitalik Buterin has pushed back against the idea that more capable AI hackers will make cybersecurity impossible to defend, arguing on Sept. 17 that advanced verification tools could eventually provide defenders with a structural advantage.
Buterin said in a Sept. 17 post on X that he disagrees with the growing belief that “AI hacking means cybersecurity is doomed.” He argued that people who continue to hold cryptocurrency are effectively betting that secure digital systems can withstand increasingly powerful automated attacks, while noting that around 90% of his own net worth is still held in crypto.
His argument is focused on formal verification, a method in which mathematical specifications and proofs are used to establish whether software operates according to predefined properties. Buterin compared it with advanced mathematical theorem proving, suggesting that sufficiently capable AI could eventually help demonstrate that “this program is secure,” even when the underlying software is highly complex.
Buterin Says Formal Verification Could Give Defenders an Edge
Buterin’s argument does not frame AI security as a race in which defensive researchers merely discover vulnerabilities before attackers do. Instead, his post outlines a model where developers specify the properties a system must meet, after which automated proof tools are used to verify that the implementation conforms to those requirements.
Ethereum’s own formal verification documentation explains the technique in more specific terms. It states that formal verification can be used to mathematically prove that a smart contract follows a formal specification, providing stronger assurances than standard testing for the properties covered by that specification.
The same Ethereum documentation highlights a key limitation relevant to Buterin’s argument. A proof can establish that software meets the specification being tested, but it does not automatically show that all required security properties were defined correctly by developers. If specifications are incomplete or poorly designed, certain behaviors can remain outside the scope of the proof.
Buterin made a similar distinction in his post. He argued that if “security” is defined too narrowly, potential attack paths involving protocols, servers, databases, networking layers, caches and other supporting components may be overlooked. His proposed approach is therefore focused on verifying more of the complete system rather than treating only a limited set of modules as security-critical.
Buterin had previously described AI-assisted formal verification as a potential “final form” of software development. In his earlier remarks, the approach was applied to areas such as Ethereum consensus, zero-knowledge systems and quantum-resistant cryptography.
Ethereum Is Already Testing AI-Assisted Security
AI is already being used by Ethereum researchers for security work, although current systems are still far from automatically proving the security of an entire system.
The Ethereum Foundation Protocol Security team reported in July that coordinated AI agents had identified real flaws in systems used by Ethereum. One confirmed issue was a remotely reachable crash in the Gossipsub networking implementation of Rust libp2p.
The vulnerability was designated CVE-2026-34219. According to the U.S. National Vulnerability Database, versions earlier than 0.49.4 could be remotely crashed by a specially crafted PRUNE message that caused an arithmetic overflow while Gossipsub backoff was being processed. The issue was fixed in version 0.49.4.
The Foundation’s security team said finding vulnerabilities was not the most difficult part of the AI-driven workflow. It noted that AI agents often produced credible reports based on unreachable execution paths, failures limited to debug environments, or formal proofs that technically succeeded while establishing a weaker property than the one originally intended.
Independent reproduction was still required before a finding could be accepted by the team. The Foundation said automated checks and human review were necessary because an agent could generate a proof that appeared valid but failed to restrict the software behavior researchers were actually seeking to verify.
The Foundation’s experience was framed as a triage challenge: large numbers of potential vulnerabilities could be generated by AI, but each issue still had to be assessed by researchers to determine whether it was reachable and meaningful in a production environment.
Formal Verification Enters Ethereum’s Protocol Roadmap
Formal verification has also been included in the Ethereum Foundation’s current protocol research program. In its Sept. 7 protocol priorities update, the Foundation said the technology would be used as cross-cutting tooling across its remaining multi-year research areas. The roadmap includes work on privacy, state, zkEVM development, post-quantum security and other protocol components through 2029.
The same roadmap states that work on an L1 zkEVM is expected to advance formal-verification tools, workflows and verified cryptographic components. Ethereum researchers are working toward a system in which validators would eventually verify concise execution proofs rather than independently re-executing each block.
A separate Ethereum Foundation project launched in August is already combining AI agents with machine-checked proofs. Through the better.codes project, researchers can direct AI systems toward a cryptographic soundness problem formalized in Lean, while submitted proofs are checked by the Lean kernel against a fixed theorem statement.
The Foundation’s Q2 funding report also includes additional projects focused on combining large language models with formal methods. These efforts include ETHeorem, which is being used to check Ethereum client implementations against specifications, along with SPECA and LeanAgent for automated protocol compliance. Further formal-verification work is also being carried out around RISC-V zkVM infrastructure.
Ethereum’s updated technical roadmap places formal verification alongside privacy, zero-knowledge proofs and post-quantum security. Buterin’s Sept. 17 remarks therefore build on an existing research direction rather than announcing a new upgrade or planned fork. No new EIP, deployment timeline or mandatory requirement for every Ethereum program to undergo formal verification was mentioned in his post.
Current AI Evidence Also Shows Stronger Offensive Capabilities
Recent cybersecurity research has reinforced concerns that AI is increasing attackers’ capabilities even as its use for defense continues to expand. In its September threat intelligence report, Anthropic said malicious actors had been observed using AI to automate vulnerability research, exploit development and campaigns targeting multiple systems. The company also reported that some operators had maintained automated workflows capable of carrying out vulnerability research on a continuous basis.
Earlier research from Anthropic had identified similar evidence on a broader scale. According to its coordinated disclosure dashboard, 2,300 vulnerabilities discovered by AI had been disclosed across 392 open-source projects by Aug. 26, with 421 of those issues having been patched upstream by that time.
Anthropic has also presented the same technology as a defensive tool. Through its Project Glasswing initiative, frontier AI models were used with security firms and software vendors to identify high- and critical-severity vulnerabilities before they could be exploited by malicious actors.
The offensive use of AI has also been documented. Anthropic’s June study examined 832 accounts linked to malicious cyber activity between March 2025 and March 2026, finding that AI was being used by threat actors across several stages of cyber operations. Buterin had previously dismissed a separate argument that AI-related security risks could seriously undermine confidence in Bitcoin. He said network-layer issues could often be resolved through software and mining infrastructure upgrades, while actual compromises of Bitcoin’s hashes or proof-of-work mechanism were described as highly unlikely.
Anthropic’s September report said its investigators had identified operators using AI-developed exploit pipelines against security appliances and government systems, with some campaigns resulting in previously unknown vulnerabilities that were validated by attackers in their own testing environments.
