Crypto's $1T defenders ask AI labs for keys to the models attackers already have

More than 40 Bitcoin and digital-asset organisations have asked leading AI labs to give vetted open-source security researchers controlled access to frontier models ^[1]. The letter, coordinated by the Bitcoin Policy Institute, presents access to systems from labs including OpenAI and Anthropic as a defensive requirement for code protecting more than $1 trillion in assets ^[2][3].

The argument is narrower than a complaint about censorship. BPI's position is that attackers can already reach advanced or locally deployed AI tools, while legitimate defenders are stopped by safety filters or access limits ^[4]. When a frontier system declines a security prompt, or puts a capability behind a programme that open-source financial infrastructure cannot join, maintainers fall back to weaker open-weight models, and attackers operate under no such constraint, according to the letter ^[6]. The signatories concede the tooling is dual-use: a model that helps a programmer find flaws in a large system helps a malicious user do the same ^[5]. What they are asking for is not fewer guardrails for everyone ^[8], but "standing trusted-access programs for qualified defenders of open-source financial infrastructure," entered through an evaluation process ^[7].

The loss data gives the request weight without settling it. SlowMist recorded 182 incidents and about $956 million in H1 2026 losses; TRM Labs counted 207 hacks and $972 million ^[9][10]. The two tallies differ by 25 incidents and roughly $16 million ^[11], which says something about how loose the measurement is, and both land near a tenth of a percent of the $1 trillion the signatories say they secure ^[12]. As an annualised loss rate on assets under custody, that is survivable. The stronger claim is about who is attacking: BPI says it has received independent reports from open-source maintainers describing sophisticated attackers using advanced AI, some possibly foreign adversaries ^[15].

Signatory quality carries this more than the headcount. Block, Coinbase, Strategy, MARA, Galaxy, BitGo, Brink, OpenSats, Chaincode Labs, Spiral, Trezor, Unchained, Btrust and Fedi are among the names reported by crypto.news ^[13], with Blockstream, Anchorage Digital, ARK Invest, Bitwise, Foundry and BTCPay Server reported by NewsBTC ^[14]. That puts exchanges, custodians, miners, corporate treasuries and the protocol maintainers themselves on one document, which is unusual and makes the letter harder for a lab's policy team to file away.

There is evidence models find real defects. An Ethereum Foundation study in July used coordinated AI agents to surface genuine vulnerabilities, including a libp2p flaw later disclosed as CVE-2026-34219 ^[21].

The awkward part is that restricted access may be a depreciating asset. Bitcoin red team contributor Calle wrote on X on 13 August that the work represented "a massive collision between decades of human open source slop against 2 weeks of kimi k3" ^[16]; Kimi comes from the Chinese lab Moonshot ^[17]. CSIS analyst Yasir Atalan said in July that Chinese models are "now months, not years" behind American ones, citing a government study that put DeepSeek V4-Pro, the best Chinese open-weight model, about eight months behind US leaders ^[18]. Stanford AI Index data published in 2026 counted 59 important US models in 2025 against China's 35, roughly 1.7 times as many, though China leads on research and patent output ^[19][20]. A defensive edge rented from US labs erodes as open-weight rivals catch up.

Watch whether a single lab actually stands up a vetting programme, who administers the vetting, and what liability the lab accepts for a credentialed researcher. Watch also whether the perimeter stays at Bitcoin, because every other open-source maintainer of financial plumbing has the same argument available.