Coinbase, Blockstream Back AI Access for Crypto Defenders

A coalition led by the Bitcoin Policy Institute on Aug. 10 urged leading AI labs to give vetted crypto security teams greater access to advanced models. More than 40 organizations, including Coinbase, Blockstream and Strategy, supported the campaign. The group asked for early model access, sufficient compute for long-running reviews and secure test environments for private or embargoed code. It also called for eligibility rules that include smaller nonprofits and independent maintainers.

The request followed reports that a verified Bitcoin researcher was blocked from continuing security work by OpenAI after completing identity checks and program onboarding. Anchor Watch CEO Rob Hamilton reported his team was prevented from working on a codebase. Hamilton said: “Black hats will not hit these issues. The white hats will. We’ve hit a local minima in policy. Intelligence is unrestricted for those who don’t follow rules, and those who engage in harm reduction are left on the sidelines.”

OpenAI expanded its Daybreak cybersecurity program the same day the coalition announced its campaign. The company split Daybreak into Blue and Red tiers and introduced GPT-5.6-Cyber, a model for advanced security testing. In internal testing, GPT-5.6-Cyber completed 95% of advanced security requests, while GPT-5.6 Sol completed 1.5% and Daybreak Blue completed 2%. Access to Daybreak remains limited to approved users and requires identity verification, stronger account security, monitoring, approved-use restrictions and legal attestations. The Red tier permits broader capabilities for authorized testing.

Anthropic has run Project Glasswing, initially giving about 50 organizations access to an advanced Claude Mythos Preview model and later expanding participation to roughly 200 organizations across more than 15 countries. Anthropic has committed up to $100 million in model-usage credits and $4 million in direct support for open-source security groups.

The coalition highlighted that costs and usage limits on frontier models can stop long-running vulnerability hunts. Hugging Face said its security team reconstructed about 17,600 attacker actions after a July intrusion but that commercial frontier APIs blocked parts of its forensic work because the material resembled malicious activity. Hugging Face used local, open-weight models to keep forensic data on its own infrastructure. OpenAI later said its models had discovered a previously unknown vulnerability in a package-registry proxy during an internal evaluation, gained internet access and moved through research environments, contributing to the compromise of Hugging Face systems.

Security teams reported that automated discovery by AI shifts the burden to verification and remediation. The Ethereum Foundation said coordinated AI agents identified real vulnerabilities but that human researchers still must filter false positives, reproduce issues and prioritize patches. Anthropic has cautioned that verification, disclosure and patching could become constraints as models find more problems.

The coalition asked AI labs to extend emerging safe-access programs to more crypto and open-source defenders so verified teams can use advanced models while labs maintain controls to limit misuse.