Forty crypto groups want the ungated model, because Coinbase already has one

On August 10 the Bitcoin Policy Institute published an open letter asking frontier AI labs to create or expand trusted-access programs so vetted security researchers protecting open-source financial infrastructure can use the labs' strongest cyber-capable models, including prerelease access where appropriate ^[1]. More than 40 Bitcoin and crypto organizations backed the request at publication, according to Decrypt reporting cited in the coverage, and their argument is about supply: safety restrictions on public models can leave defenders holding weaker tools than attackers may obtain elsewhere ^[2][3][11].

The ask is more specific than a statement of principle. Alongside model access, the letter requests sufficient compute, long-horizon agent use, secure environments for reviewing private code, eligibility for small and independent maintainers, and a direct disclosure channel to model-lab security teams ^[4]. That last item is the sort of thing you only write down after trying to report a finding and discovering there is no address to send it to. BPI's technical case is that advanced models help researchers inspect large codebases, identify potential vulnerabilities and accelerate difficult security work, and that the same capabilities help attackers ^[5][6]. Nobody in this argument disputes the dual use. The dispute is over who gets on the list.

Selective admission is already happening. Cointelegraph reported on August 3 that Coinbase obtained access to Anthropic's restricted Mythos model in June ^[7], and that Mythos 5 uses the same underlying model as the publicly available Fable 5, without the safeguards that restrict sensitive cybersecurity work ^[8]. The same outlet reported that Zooko Wilcox said Anthropic used the model to audit the Zcash protocol at Shielded Labs' request ^[9]. Other firms did not have comparable access at the time ^[10]. Binance Chief Security Officer Jimmy Su told Cointelegraph the exchange had sought access to more frontier models and had not received Mythos, and also called a limited initial rollout responsible when a model could improve attacker capabilities faster than defender capabilities ^[12][13]. The excluded party endorsing the gate is the most interesting line in the file.

Tiering is not unique to Anthropic. Cointelegraph reported that OpenAI gives verified defenders access to GPT-5.5 through its Trusted Access for Cyber program while reserving the more permissive GPT-5.5-Cyber for a smaller group performing authorized penetration tests ^[14]. So the machinery exists at more than one lab. What does not exist publicly is the admission criteria, which is why a letter signed by 40-plus organizations arrived seven days after the report describing who already got in ^[15].

Two things to watch. First, whether any lab publishes eligibility rules, and specifically whether they can admit a two-person library maintainer, since vetting regimes built around corporate counterparties tend to screen out exactly the unfunded maintainers holding up open-source financial plumbing ^[4]. Second, whether the firms already inside publish findings, because an audit that produces nothing citable is not evidence that access works. This is a proposal, not a confirmed change in access policy, and the retrieved reporting identified no public response from the major labs ^[16][17].