Bitcoin Red Team is forced to return to Chinese AI models due to OpenAI blocks.
The situation surrounding access to advanced AI tools for cybersecurity is taking an alarming turn. Rob Hamilton, head of AnchorWatch and an active participant in the Bitcoin Red Team initiative, has encountered an unexpected obstacle from OpenAI. After starting the integration of specialized cyber models into the audit process on August 8, he discovered the very next morning that access to them had been restricted.
Hamilton emphasizes that the block occurred after fully completing KYC and onboarding procedures, making the situation particularly telling. As a result, he had to urgently switch to open Chinese models, which, in his words, are becoming the only viable alternative for ecosystem defenders.
This creates a dangerous imbalance. While legitimate security researchers face bureaucratic barriers and restrictions, malicious actors are clearly not bound by such obligations. Hamilton states outright: this approach by American AI giants cuts defenders off from the most powerful tools, effectively playing into the hands of hackers.
The situation with Anthropic appears to be no better. Hamilton notes that although he has not yet encountered problems, his access will likely be downgraded from the Fable tier to Opus, which is essentially the same restriction in a different form.
For context, Bitcoin Red Team is a volunteer group that uses AI and manual review to analyze hundreds of open-source repositories related to Bitcoin. To date, the team of 25 people has identified 1,288 critical and high-level vulnerabilities in the ecosystem. This is serious work, and its effectiveness directly depends on access to the best AI models.
Notably, despite the blocks, the team's AI spending continues to grow. Total costs have already exceeded $58,000. Moreover, although OpenAI and Anthropic cyber models show impressive results, the bulk of the budget (nearly 75%) still goes to the Chinese model Kimi K3. The share of Qwen 3.8 from Alibaba is also growing. As Hamilton ironically noted, for him and his colleague calle, a "very Chinese period of life" is beginning.
My analysis: This situation is a vivid example of how geopolitical restrictions in the AI field directly damage the security of critical infrastructure. While regulators and corporations debate policy, real defenders are forced to seek workarounds, and Chinese models are becoming the de facto standard for security auditing. This is not just an inconvenience but a systemic risk that requires a reassessment of approaches to export control in the AI domain.