How Bitcoin is hardening itself in 2026.

The big losses in 2026 came from a wallet's firmware, a payment server and an exchange. Bitcoin's own rules and signatures held up. What I wanted to look at is what the people who maintain Bitcoin are doing about two things at once: AI making bugs cheap to find, and a quantum timeline that keeps getting pulled in.

The response has been to point the same AI tools at the code before attackers do.

• Project Loupe. Spiral and Block launched a free, open-source AI scanner for Bitcoin projects in May. It only reports findings that come with a working test case. The trial run across Bitcoin Core, rust-bitcoin, LDK, BDK and others turned up 643 findings and more than 70 fixes.
• The Bitcoin Red Team. Sixteen volunteer developers got together after the Coldcard losses, ran AI models across 390 open-source Bitcoin projects in about a day and logged 4,962 potential flaws (85 rated critical). Those are raw scan results. Nobody has published how many hold up.
• OpenSats Red Team Fund. Pays the AI and infrastructure costs of finding and disclosing bugs in critical Bitcoin software. The first grants went out October 12 to work on Bitcoin Core, libsecp256k1, mempool.space, btcd and lnd, and BTCPay.
• Brink. About $7.8 million in 2025 donations paid for eight engineers, Bitcoin Core's first public third-party audit (by Quarkslab) and Fuzzamoto, a framework that tests implementations against each other. A bug bounty is on their 2026 list.
• Fuzzing. Bitcoin Core has run more than a century of CPU time feeding junk inputs to its own code.

One Brink-funded engineer says he's seen more than 1,000 LLM-generated reports against Bitcoin Core. Most were false positives or hypothetical internal issues, and none came in at high or critical. He says himself that he's biased.

Bitcoin's signatures use elliptic-curve cryptography, and a big enough quantum computer could break it. In March, Google Quantum AI estimated it could take fewer than 500,000 physical qubits to break a 256-bit elliptic-curve key, about 20 times fewer than earlier estimates. Google has set 2029 for its own migration, and NIST guidance points to phasing out ECDSA and RSA by 2030.

The hardware is a long way off. The biggest public quantum attack on an elliptic-curve key broke 15 bits in April (Bitcoin uses 256). Chaincode Labs sees it as roughly a two-year emergency track and a seven-year full transition.

Where it actually gets decided is the proposals, and they're being argued in the open:

• BIP-360 (Pay-to-Merkle-Root) went into the BIP repository as a draft in February. It removes the quantum-exposed key-path spend from a Taproot-style output. It doesn't add post-quantum signatures on its own.
• BIP-361 proposes phasing out legacy signatures, which would eventually freeze coins that never move. One of its authors calls it "a rough sketch" and critics call it confiscation. Adam Back wants optional upgrades now and fast coordination later.
• SHRINCS. Blockstream Research's hash-based signature work only relies on assumptions Bitcoin already trusts. There's a draft spec. The standard NIST schemes would cut throughput a lot and SHRINCS narrows that gap, but a formal security proof is still to come.

There's no agreed post-quantum signature scheme yet and no migration date. About a third of supply sits in outputs whose public keys are already visible on-chain, including the very early coins and reused addresses. Coins in hashed addresses that have never been spent from haven't shown their keys.