A team of sixteen volunteers filed 4,962 vulnerability reports across 390 Bitcoin ecosystem projects during a flash audit lasting roughly 30 hours in early August 2026, relying primarily on AI agents. The operation, funded by OpenSats for about $40,000, signals a turning point in how open source security reviews are conducted at scale.
🔑 Key takeaways
- 4,962 findings filed across 390 projects in ~30 hours by the Bitcoin Red Team
- 720 critical or high-severity vulnerabilities (14.5% of total), averaging 1.85 serious issues per project
- 91% of discoveries generated through automated AI ingestion, supervised by 14 humans and 3 automated agents
- Funding of approximately $40,000 provided by the OpenSats foundation
- Bitcoin core protocol was not audited: only user-facing tools (wallets, SDKs, exchanges) were in scope
An industrial-scale flash audit
The Bitcoin Red Team worked continuously for around thirty hours in early August 2026, drawing on sixteen operators distributed across multiple time zones. According to a situation report published by calle, creator of the Cashu Bitcoin ecash protocol, the final roster brought together 17 contributors: 14 humans and 3 automated agents. The sustained cadence — between 166 and 180 reports per hour — is rarely matched by traditional audit cycles.
Funding, supplied by OpenSats (a non-profit that backs Bitcoin open source development), totaled roughly $40,000. The envelope covered compute infrastructure and coordination, without any compensation paid to the contributors, who all worked on a volunteer basis.
« We are sincerely sorry if our reports added stress to an already stressful day. »
calle, Cashu developer and Bitcoin Red Team coordinator
Severity concentrated in select segments
While the total volume of reports is sizable, severity is concentrated in specific project categories. Of the 4,962 reports, 85 were classified as critical and 635 as high severity, together accounting for 14.5% of the corpus. The average stands at 1.85 serious issues per audited project, an unprecedented figure at this scale.
Approximately 91% of discoveries were generated through automated ingestion, but calle stressed that human supervision remained essential to filter false positives and validate relevance. 21% of reports were accompanied by dynamic reproduction, i.e., proof-of-concept code demonstrating the vulnerability.
| Project category | Total reports | Critical + high share |
|---|---|---|
| Privacy and coinjoin tools | ~620 | 24% |
| Exchange and swap platforms | ~770 | 21% |
| Payment tools and merchant solutions | ~720 | 17% |
| Crypto libraries and SDKs | 1,101 | 10% |
Crypto libraries and software development kits produced the largest raw volume (1,101 reports), but only one in ten reached the high-severity threshold. By contrast, privacy and coinjoin tools, whose attack surface is typically more complex, concentrate the highest share of serious vulnerabilities (24%).
The Coldcard affair as a catalyst
The campaign was launched a few days after the public disclosure of a flaw in the Coldcard hardware wallet from Coinkite. The defect, present since March 2021 in a firmware build, extracted wallet seeds from a software pseudo-random number generator rather than from the device’s hardware TRNG (True Random Number Generator), making private keys guessable.
In its post-mortem, Coinkite estimated that « someone probably used AI to review earlier firmware versions ». The incident reportedly cost users about $130 million in Bitcoin, underscoring the magnitude of consequences when a hardware vulnerability remains undetected for over five years.
« AI is now used to find vulnerabilities in crypto code at machine speed. Open source and reviewed don’t mean the same thing. »
Charles Guillemet, CTO at Ledger
Rapid disclosure fuels tension
At the time the report was published, only 19 projects — less than 5% of those audited — had received prior notification. Calle acknowledged the additional burden placed on maintainers, while defending the rapid disclosure pace: project owners are best placed to validate findings, validation has become « nearly free thanks to AI, » and anyone using the same tools would reach the same bugs. Eight reports have already been withdrawn as false positives.
Prior to the campaign, the team had already scanned roughly 150 repositories, leading to more than a dozen private disclosures. For cases deemed critical, vulnerabilities were reproduced locally before maintainers were contacted, in a responsible-disclosure approach. Contributors include Rob Hamilton, CEO of AnchorWatch, and developer calle (Cashu).
Conclusion: AI reshuffles the open source audit deck
This campaign demonstrates that combining AI agents with distributed human coordination can produce in a few hours an audit volume that traditional teams would need months to match. It also confirms a point raised by Charles Guillemet: open source code is not equivalent to thoroughly reviewed code. The Coldcard flaw, hidden for more than five years in public firmware, stands as the costliest example of the year.
The Bitcoin Red Team has announced plans to release its tooling as open source, which could accelerate the standardization of such approaches across the broader crypto ecosystem. The scenario to watch: a growing share of maintainers integrating AI agents into their CI/CD pipelines, turning spot security audits into a continuous process. Conversely, a wave of mass disclosures without sufficient coordination could overwhelm small teams and, paradoxically, slow the rate at which patches ship.
Sources
This article is published for informational and educational purposes only. It does not constitute investment advice. Do your own research (DYOR) before making any decision.
