Bitcoin Companies Call on AI Labs: Defenders Lag Behind Attackers

Bitcoin and crypto companies have called on labs to provide open-source security researchers with early access to the most powerful AI models; the request is based on the argument that defenders are working with weaker tools than attackers.

Organized by the Bitcoin Policy Institute and published this week, the letter bears the signatures of companies such as Coinbase, Block, BitGo, Blockstream, Anchorage Digital, ARK Invest, Bitwise, Foundry, Casa, and Exodus, as well as developer funds like Brink, Chaincode, and Btrust.

The letter states that Bitcoin Core developers, who maintain the software that runs the network, cannot access the programs that AI labs offer to their security partners. While malware filters in public models also hinder vulnerability discovery efforts, teams are forced to settle for lower-capacity, downloadable models. While labs and a limited number of partners can see new attack capabilities months in advance, attackers are not subject to these restrictions.

Crypto companies want early access to AI models

The signatories list five demands: early access to the strongest cybersecurity models, a sufficient compute budget for meaningful reviews, a secure environment where private code can be inspected, the inclusion of small and independent developers in the program, and direct communication with lab security teams to convey findings.

The urgency behind the request has become apparent through events in recent weeks. BTCPay Server, a signatory of the letter, disclosed a critical vulnerability last week used by attackers that led to the draining of merchants’ Lightning nodes. Another signatory, Foundation, lost its own node in the same attack.

In response, the Bitcoin Red Team reported thousands of findings across hundreds of projects this month by using AI models on Bitcoin codebases, one of which prompted BTCPay Server to release a patch. While the request has yet to be met, there is no evidence that the attack in question was carried out using artificial intelligence.