Vitalik Buterin Rejects ‘AI Hacking Means Cybersecurity Is Doomed’ — and Points to Formal Verification as the Defense
Ethereum co-founder Vitalik Buterin has publicly rejected one of the most common doom scenarios in modern security discourse: the idea that increasingly capable AI hackers will eventually make cybersecurity unwinnable. In a September 17 post on X, Buterin argued the opposite — that cybersecurity is “naturally defense-favoring once people get their shit together,” and that advanced verification tooling could give defenders a structural edge for the first time.
Buterin put his money where the argument is. Anyone who continues holding cryptocurrency, he wrote, is implicitly betting that secure digital systems can survive stronger automated attacks — and roughly 90 percent of his own net worth remains in crypto.
Formal verification as the counterweight
The technical core of Buterin’s argument is formal verification: the use of mathematical specifications and machine-checked proofs to establish that software behaves according to predefined properties. Using advanced theorem proving as an analogy, he suggested that sufficiently capable AI could help prove that “this program is secure” even when the software is highly complex.
Ethereum’s own documentation describes the technique in narrower terms, noting that formal verification can mathematically prove a smart contract complies with a formal specification — stronger guarantees than testing, but only for the properties included in the specification. Buterin acknowledged the limitation directly in his post: defining “security” too narrowly can omit attack paths involving protocols, servers, databases, networking layers, caches and other supporting components. His proposed direction is to verify more of the complete system rather than labeling a small group of modules security-critical.
The comments extend a position Buterin has staked out before, having previously described AI-assisted formal verification as a possible “final form” of software development with applications across Ethereum consensus, zero-knowledge systems and quantum-resistant cryptography.
Ethereum is already running the experiment
The argument is not hypothetical for the Ethereum ecosystem. The Ethereum Foundation’s Protocol Security team reported in July that coordinated AI agents had found real defects in systems Ethereum depends on — including a remotely reachable crash in Rust libp2p’s Gossipsub implementation, catalogued as CVE-2026-34219. Versions before 0.49.4 could be crashed remotely through a crafted PRUNE message that triggered an arithmetic overflow during backoff handling; the fix shipped in 0.49.4.
The team’s most important finding, however, was about the workflow, not the bugs. Researchers said vulnerability discovery was not the hardest part: AI agents frequently generated convincing reports involving unreachable execution paths, debug-only failures, or formal proofs that technically passed while proving a weaker property than intended. Independent reproduction remained mandatory before any finding was accepted, because an agent can produce a valid-looking proof that fails to constrain the behavior researchers actually wanted to test. The Foundation described the challenge as a triage problem — separating reachable, meaningful vulnerabilities from an avalanche of plausible-looking candidates.
Formal verification is also moving into Ethereum’s official roadmap. In its September 7 protocol priorities update, the Foundation designated formal verification as cross-cutting tooling across its remaining multi-year research areas, including privacy, state growth, zkEVM development and post-quantum security through 2029. The L1 zkEVM workstream is expected to advance verified cryptographic components as validators move toward verifying succinct execution proofs instead of re-executing every block.
In August, the Foundation launched better.codes, a project combining AI agents with machine-checked proofs: researchers aim AI at cryptographic soundness problems formalized in Lean, while the Lean kernel checks whether submitted proofs satisfy a fixed theorem statement. The Q2 funding report lists further LLM-plus-formal-methods work, including ETTheorem for checking Ethereum client implementations against specifications, SPECA and LeanAgent for automated protocol compliance, and formal verification of RISC-V zkVM infrastructure.
The attackers are not standing still
Buterin’s optimism comes against a backdrop of evidence that AI is strengthening offense as well. Anthropic’s September threat intelligence report said frontier AI can accelerate attacks while its models have identified thousands of flaws in real software. The open question is which side compounds faster.
Buterin’s post did not name a new EIP, deployment date or mandatory verification rule — it is a thesis, not an upgrade. But with 90 percent of his net worth riding on the answer, the bet is now explicit: if provable software beats automated exploitation, the holders win. If not, no amount of net worth will buffer the outcome.
At press time, Ethereum was trading around 2,459 USD, Bitcoin near 76,538 USD and Solana around 101 USD, according to CoinGecko data.
90 percent of his net worth in crypto while arguing security is winnable. at least the man’s consistent
skin in the game argument only goes so far. plenty of founders rode their bags to zero fully convinced they were right
true but the 90 percent part is just context. the actual claim is defense favoring once tooling ships, and lean style provers are getting there faster than anyone expected
formal verification is great until the spec itself is wrong. proving code matches a bad spec just gives you confident bugs
“confident bugs” is a great phrase and yes thats the real risk. but spec review is a human problem, the proving half gets cheaper every year
^ this is the actual bottleneck. spec writing is the hard part, the proving gets easier every year
spec gaps are real but thats exactly what modern spec languages attack now. we went from hoping the code matches intent to proving it, the gap shrinks every year
confident bugs, exactly. the fix isnt better proofs tho, its spec review by humans who didnt write the code
fair, but his answer to that in the thread was you verify the spec against requirements too. at some point you are verifying the verifier, which sounds absurd until you compare it to teams shipping unaudited contracts today
verifying the verifier sounds absurd until you remember thats literally how coq and lean ecosystems already work. the tower of proofs bottoms out somewhere small enough to actually read
coq kernels are tiny and battle tested for decades. the trust anchor is way smaller than any audited contract, thats the whole trick
confident bugs is only fatal if nobody reads the spec. formal methods people have been fighting spec rot for decades, tooling helps there too
90% of his net worth still in crypto is honestly the strongest part of the argument. dude is betting his own money on verification tooling working
“once people get their shit together” is carrying that whole argument lol. most teams still ship unaudited contracts and always will
most teams always will ship unaudited, but the point is tooling becoming default so lazy teams inherit some safety anyway. nobody handrolls tls anymore either
counterpoint: formal methods went from academic curiosity to usable tooling in like five years. the getting shit together part is closer than you think
people said the same about fuzzing until it shipped default in every ci pipeline. tooling normalizes faster than anyone predicts
skin in the game is nice, but Vitalik also admitted ethereum culture undervalues security until something breaks. hope the audit grants match the tweets
audit grants fix yesterday’s bugs. verification tooling shipping in default toolchains fixes tomorrow’s, and that is the actual gap he is pointing at
90 percent of net worth in crypto is less a bet on verification and more just refusal to diversify lol. respect either way
specification gaps are the real problem. proving a contract secure means nothing if the spec forgets the oracle layer or the servers around it
he literally addressed that in the post, verify more of the complete system instead of cherry picking modules. read past the headline
defenders have been structurally behind since the 90s. if verification tooling actually flips that, first defense-favoring era in security history, thats a bigger deal than people realize
spec bugs are real but the alternative is pray and ship. ill take provable code with a human reviewed spec over vibes any day
formal verification lands on chain contracts way before web2 ever touches it. the money at risk forces the tooling first