Crypto Firms Seek Frontier AI Security Access
More than 40 Bitcoin and crypto firms asked frontier AI labs this week to give vetted security researchers access to their strongest models before public release, seeking better tools to defend open-source financial infrastructure. The Bitcoin Policy Institute's open letter says restrictions can leave defenders with weaker tools than attackers may obtain. Coinbase has received access to Anthropic's restricted Mythos model, according to Cointelegraph.
More than 40 Bitcoin and crypto firms have asked leading AI labs to create or expand trusted-access programs for security researchers protecting open-source financial infrastructure, Decrypt reports. The request seeks access to powerful cyber-capable models before they are broadly released, subject to vetting for qualified defenders.
According to crypto.news, the Bitcoin Policy Institute (BPI) published an open letter calling for standing access programs that would let approved maintainers and researchers use capabilities otherwise restricted by model safeguards or specialized cyber programs. Anchorage Digital, BitGo, Bitwise, Blockstream, Kraken, Ledger, MARA and Trezor were among the listed signatories, crypto.news reports.
BPI argued that advanced models can help researchers inspect large codebases, identify potential vulnerabilities and accelerate difficult security work. The same capabilities can aid attackers, creating the core policy tension behind controlled access to frontier cyber models.
Uneven access to restricted models
Cointelegraph reports that Coinbase obtained access to Anthropic's restricted Mythos model in June. It also reported that Zooko Wilcox said Anthropic used the model to audit the Zcash protocol at Shielded Labs' request.
Other firms remain without comparable access. Binance Chief Security Officer Jimmy Su told Cointelegraph that Binance had sought access to more frontier models but had not received Mythos. Su also described a limited initial rollout as responsible when a model could improve attacker capabilities faster than defender capabilities.
According to Cointelegraph, Anthropic's Mythos 5 uses the same underlying model as publicly available Fable 5, without safeguards restricting sensitive cybersecurity work. The outlet also reports that OpenAI offers verified defenders access to GPT-5.5 through its Trusted Access for Cyber program, while reserving its more permissive GPT-5.5-Cyber model for a smaller group performing authorized penetration tests.
Security-access tradeoff
The letter centers on a practical asymmetry: public model safety controls can block legitimate vulnerability research, while attackers may use less restricted tools or independently available models. BPI's position, as reported by crypto.news, is that qualified defenders of open-source financial systems should be included in trusted-access programs.
The crypto firms' request places Bitcoin and related open-source financial software directly into the broader debate over whether frontier-model cyber controls can distinguish authorized defensive research from harmful use.
Key Points
- 1More than 40 crypto firms requested vetted frontier-model access, framing code-security research as an urgent defensive use case.
- 2Cointelegraph reports uneven access already exists, with Coinbase and a Zcash audit receiving Anthropic Mythos model access.
- 3The reporting highlights a security tradeoff between restricting cyber-capable models and giving qualified defenders access to advanced tools.
Scoring Rationale
The request concerns access governance for cyber-capable AI models and the security of high-value open-source financial infrastructure. It is notable for AI security practitioners because it highlights the practical tension between model safeguards and authorized defensive research, although it is not a model release or confirmed policy change.
Sources
Public references used for this report.
Practice with real FinTech & Trading data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all FinTech & Trading problems
