Buterin: AI is unlikely to break Bitcoin’s math, but the practical risks are real
A public exchange between investor Liron Shapira and Ethereum co‑founder Vitalik Buterin replied the same day clarified an important distinction: advancing AI changes where attackers will strike, not necessarily the cryptographic foundation Bitcoin rests on.
On Sept. 7 Liron Shapira assigned a “50% confidence” to his forecast that AI could erode confidence in Bitcoin’s security and cause the price to fall by more than 50% within two years. Vitalik Buterin replied the same day that he “took the opposite side, ” calling the probability of “actual breaks on hashes or PoW” “tiny.” These competing judgments matter less for predicting price than for deciding where to focus defensive effort.
Two risk categories every executive should separate
Think of the threat landscape as two buckets:
- Operational infrastructure risks: wallets, exchange custody, node software, mining pools, network routing and human procedures. These are the systems attackers target first.
- Cryptographic primitive risks: the math, SHA‑256 hashing used in proof‑of‑work and elliptic‑curve signatures that protect private keys. A true cryptographic collapse would be existential, but it is a different and much harder problem than operational compromise.
History and recent incidents show attackers exploit the outer layer long before they break the math. That is the practical lesson behind a July incident reported by Bitcoin Optech. Some Coldcard‑generated wallets used insufficient entropy during key generation, and a technical report put estimated losses at more than 1, 000 BTC. The ledger itself and SHA‑256 remained intact. The loss came from tooling and process failures.
Why AI isn’t an instant route to cracking SHA‑256
SHA‑256’s security is based on preimage and collision resistance: finding an input that produces a chosen hash is effectively an exhaustive search across an enormous space. Machine learning excels at pattern recognition and approximation, not turning an exponential search into a polynomial one. Training a model to “guess” SHA‑256 preimages doesn’t bypass the resource limits that make brute force infeasible.
Quantum computing presents a separate theoretical threat to elliptic‑curve signatures (via Shor’s algorithm) and would require different timelines, resources and remediation paths than the AI‑driven operational risks discussed here. Conflating them mixes distinct problems and policy responses.
AI widens the attack surface by accelerating existing vectors
Advanced AI amplifies both offense and defense. The near‑term threats are practical and operational:
- AI‑generated spear‑phishing that personalizes messages from a user’s wallet provider or exchange support.
- Automated code‑audit tools that quickly surface zero‑day bugs in wallet firmware or node clients.
- Malware that harvests mnemonic phrases, captures screens, or bypasses multi‑factor controls using social engineering at scale.
- Attacks on mining‑pool software, routing, or exchange custody processes that cause liquidity stress or service outages.
Defenders can use AI for rapid code review, anomaly detection and incident triage, but only if they have comparable access to frontier models, test environments and compute resources. That is the core of the “transition” concern raised by researchers such as Zack Shapiro and Efrat Fenigson and by the Bitcoin Policy Institute. If offensive actors get advanced AI first, defenders may be playing catch‑up.
Industry asks: controlled access for defensive research
In August, more than 40 digital‑asset organizations appealed to leading AI labs to grant vetted open‑source security teams controlled access to advanced models and testbeds. Signatories included Block, Coinbase, Marathon Digital Holdings (MARA), Galaxy, BitGo, Brink and Trezor, along with several Bitcoin development groups. The goal is simple: reduce asymmetric access to offensive capabilities so defenders can harden systems before exploits are weaponized.
What would cause major market moves?
A large, visible operational breach, a custody failure, widespread wallet theft, or coordinated attack on exchanges, could trigger a sudden loss of confidence and forced selling, which would move price. That is a plausible channel for a sharp decline and the type of outcome Shapira was warning about. It is different from an immediate cryptographic collapse of SHA‑256 or proof‑of‑work, which Buterin judges unlikely on a two‑year horizon.
Concrete checklist for custodians, exchanges and boards
Executives need clear, prioritized actions. These steps are tactical, achievable, and map directly to the operational risks AI amplifies.
- 30 days, Entropy & key hygiene audit: Verify randomness sources for key generation. Adopt hardware RNGs with attestations, run statistical tests per NIST SP 800‑90 guidance, and publish an audit trail for key‑generation procedures. If you use vendor devices like Coldcard, follow vendor advisories and forensics reports.
- 60 days, Patch automation and hardening: Inventory critical clients, node software, and mining‑pool components. Automate critical security patch deployment, require signed releases, and enforce code‑signing for production firmware.
- 90 days, Threat modeling & tabletop drills: Run scenario exercises that include AI‑scaled phishing, coordinated custody compromise, and supply‑chain attacks. Validate recovery plans and withdrawal/transfer limits under stress.
- Ongoing, Vet frontier AI access: Support or join controlled‑access programs so vetted security researchers can test models defensively. Implement policy controls (dual‑control, audit logs) before any internal model testing touches production secrets.
- Governance, Communication & migration planning: Build a cross‑stakeholder governance playbook (nodes, miners, exchanges, custodians) describing how you would coordinate if a deep cryptographic fix became necessary, for example a coordinated signature migration to post‑quantum algorithms.
Sources and suggested reading
Primary material worth reviewing for technical context and original wording:
- Liron Shapira’s Sept. 7 public forecast assigning “50% confidence” to a >50% Bitcoin decline.
- Vitalik Buterin’s Sept. 7 responses rejecting a near‑term cryptographic collapse and calling the probability of PoW or hash breaks “tiny.”
- Bitcoin Optech’s July report on the Coldcard wallet entropy issue and the accompanying technical forensic report estimating losses at more than 1, 000 BTC.
- The August open letter from 40+ digital‑asset organizations requesting controlled access for vetted security teams to frontier AI models (signatories include Block, Coinbase, MARA, Galaxy, BitGo, Brink and Trezor).
Key takeaways: questions a curious leader would ask
-
Is a direct AI‑driven break of Bitcoin’s proof‑of‑work likely within two years?
Vitalik Buterin calls that probability “tiny.” Many cryptographers and practitioners treat a direct SHA‑256 break as unlikely in the near term; the more realistic threat is AI‑amplified operational compromise.
-
Have real losses already occurred because of non‑cryptographic failures?
Yes. Bitcoin Optech reported a July Coldcard key‑generation entropy failure; a technical forensic report estimated losses exceeding 1, 000 BTC. That incident shows infrastructure mistakes, not broken hashes, cause large damage.
-
Could AI give attackers an edge over defenders?
Yes. Researchers and the Bitcoin Policy Institute warn that a transition window where offensive actors access advanced models before defenders could create dangerous asymmetries. Controlled access for vetted defensive teams is a proposed mitigation.
-
What should defenders do right now?
Start with entropy audits, hardened client/firmware practices, automated patching, tabletop exercises for AI‑scaled attacks, and participation in vetted model‑access programs. These steps close the most likely attack vectors without touching Bitcoin’s core protocol.
AI reframes the race around Bitcoin: it speeds both attack and defense, but it doesn’t hand attackers a simple shortcut into SHA‑256. The immediate job for executives and security teams is concrete and operational, harden the tooling and the human processes attackers will try to exploit first, and push for parity of access so defenders aren’t left behind.