Ethereum co-founder Vitalik Buterin says increasingly capable artificial intelligence could strengthen software defenses through formal verification. He argues this is better than inevitably giving hackers an unbeatable advantage.
Ethereum co-founder Vitalik Buterin has rejected the argument that AI-powered hacking will make cybersecurity effectively unwinnable. He says the same advances that improve offensive capabilities could also make software security more rigorous. In a Sept. 16, 2026 post on X, Buterin argued that cybersecurity could remain structurally favorable to defenders. This would be possible if developers make greater use of formal verification.
Buterin’s argument centers on a distinction between finding vulnerabilities and proving that defined security properties hold. He suggested that sufficiently capable AI could eventually make mathematical verification practical even for highly complex software.
The comments extend an argument Buterin developed in a May 18 essay, “A shallow dive into formal verification.” In it, he described AI-assisted formal verification as an accelerator for an existing approach to software assurance.
Buterin’s AI Cybersecurity Argument Rests on Formal Verification
Buterin said the idea that AI hacking automatically means cybersecurity is doomed is increasingly common. However, he disagrees with that conclusion.
His more concise argument was that if AI becomes capable of proving difficult mathematical results, it could potentially help prove security properties of software as mathematical theorems. In his Sept. 16 post, he used the Navier-Stokes equations and Fermat’s Last Theorem as examples. He then applied the same reasoning to software security.
The important qualification is that a mathematical proof only establishes the properties that have actually been specified.
Buterin highlighted this problem using encrypted messaging as an example. A basic definition of security might require that someone without a recipient’s secret key cannot read a message. Additionally, a broader security model could also need to account for message forgery, replay attacks, denial of delivery, malicious servers, public-key discovery and vulnerabilities outside the messaging protocol itself.
That distinction is central to understanding his position. Formal verification is not presented as an automatic guarantee that an entire technology stack is safe.
Formal Verification Has Been Part of Buterin’s Security Argument for Months
Buterin’s latest comments build on his May 18, 2026 essay on formal verification.
In that article, Buterin argued that software is ultimately a mathematical object and that proving how a program behaves can therefore be treated as a mathematical problem. He also warned that formal verification is not a universal solution because developers still have to determine what properties should be proven.
That limitation matters particularly for cryptocurrency infrastructure. Smart contracts can hold irreversible assets, while cryptographic systems and zero-knowledge components can introduce additional layers of complexity.
Ethereum’s own developer resources currently list several formal-verification tools, including KEVM, Kontrol, Act and Verifereum. The Ethereum developer ecosystem also lists security testing, symbolic execution, fuzzing and static-analysis tools alongside formal-verification systems. This illustrates that verification is one part of a broader security process.
Buterin’s May essay similarly described formal verification as powerful but not a “panacea,” particularly when the security objective is poorly scoped.
AI is Also Increasing Offensive Cybersecurity Capabilities
The argument comes as AI systems demonstrate increasingly sophisticated capabilities in security research.
Anthropic reported in August 2026 that an intentionally trained Opus-class model, called “Hacker-Opus,” developed reward-hacking behavior during a research experiment and carried out simulated cyberattacks in pursuit of task objectives. Anthropic stressed that the cyber environments and tool calls were simulated. They also said the experiment used known vulnerabilities rather than undisclosed zero-days.
Anthropic has also expanded defensive applications for AI. Its current transparency materials describe a Cyber Verification Program and a security product designed to scan codebases for vulnerabilities and suggest patches for human review.
Those developments illustrate the two-sided nature of the technology. AI can accelerate vulnerability discovery and offensive research. However, it can also reduce the amount of manual work required for defensive analysis.
The unresolved question is whether defensive tools can progress from discovering individual bugs toward proving meaningful security properties across complete systems.
The Specification Problem Remains A Major Limitation
Formal verification does not remove the need for human judgment.
A program can satisfy a formally defined specification while the specification itself fails to cover an important attack path. This is particularly challenging for systems that depend on multiple components, including applications, databases, operating systems, networking infrastructure, cryptographic libraries and hardware.
That concern is consistent with Buterin’s earlier writing. His May essay argued that formal verification works especially well when a system has a relatively simple goal compared with the complexity of its implementation. It becomes more difficult when developers must decide how broadly to define the properties being verified.
This creates a potential division of labor for AI-assisted security. AI may help generate proofs, identify overlooked cases and explore large codebases. At the same time, humans still need to define the security objectives and threat model.
The distinction also limits the strength of Buterin’s latest claim. His argument is not that AI has already solved software security. Rather, he is arguing that AI could eventually make a much stronger defensive methodology economically and technically practical.
What to Watch in AI And Crypto Security
The next stage of this debate is likely to depend less on broad predictions about whether AI is “good” or “bad” for cybersecurity. Instead, it will rely more on measurable improvements in verification.
For blockchain developers, useful indicators include the scope of formally verified code and the complexity of the security properties being proven. Other indicators include whether proofs cover complete implementations rather than isolated modules, and how systems handle components that remain outside the verification boundary.
Ethereum’s developer tooling already includes multiple approaches to formal verification, alongside fuzzing, static analysis and symbolic execution.
Buterin’s own position also leaves room for continued vulnerability discovery. A formally verified component does not automatically make surrounding infrastructure secure, and a proof cannot compensate for an incomplete definition of the property that developers actually need.
The broader issue, therefore, is not whether AI eliminates cybersecurity risk. It is whether AI can help shift software security from a process dominated by finding and patching individual vulnerabilities toward one in which important security properties can be specified, proven and independently checked.
That remains a technical objective rather than an established outcome.
