More than 40 Bitcoin and encryption companies, including Coinbase, Block and BitGo, have called for large AI laboratories to provide more robust models of credible access to qualified independent security researchers before the model is publicly released, so that key gaps can be identified and addressed earlier.
Joint letters point to open source security.
The open letter was issued on 10 August and was organized by the Bitcoin Institute of Policy and was signed by Blockstream and ARK Invest. According to the United States, the Bitcoin network protects more than $1 trillion in value and the wider encryption infrastructure carries trillion in assets.
The letter referred to the extensive reliance on open-source codes for wallets, signature equipment, software repositories and hosting systems, and a loophole could result in loss of user assets. However, many of those who are truly responsible for the maintenance of these systems still do not have access to the access project set up by the AI Laboratory for the examined researchers.
Maintainer says it's hard to use the strongest model.
The co-sponsors indicated that, if changed to a publicly available model, the associated security restrictions often blocked normal gap testing. As a result, front-line maintainers rely more on self-deployable open-source weight models, but there is still a gap between such tools and front-line models.
The open letter stated that if earlier access and a reasonable budget were available, it would make it easier for core developers to identify and repair loopholes before losses occurred. It was also noted that the ability to attack would normally flow between the laboratory and a few partners and then spread to a wider range, leaving the defensive side at a disadvantage.
15 million dollars in previous commitments
At the end of last month, Bitcoin Security Consortium has committed $15 million for the long-term security of the Bitcoin network. Participants included BlackRock, Coinbase, Strategy, Anchorage, Ark, Blockstream, Fidelity Digital Assembly and Galaxy.
However, this joint letter emphasizes not long-term funding, but currently available modelling capacity. Anthropic had previously demonstrated that similar models could be used to launch a real-time bitcoin extortion attack with AI and that such tools were reducing the cost of the attack. As a result, the United Nations requested that the AI Laboratory expand the coverage of its credible access programme.