Bitcoin Red Team finds 85 critical flaws across 390 open-source repos

The Bitcoin Red Team filed 4,962 findings, including 85 critical flaws, across 390 open-source repos after the Coldcard exploit drained over $100 million.

Galaxy Research head of research Alex Thorn said at least 15 separate attackers exploited the Coldcard vulnerability, with confirmed losses reaching 1,596 BTC from roughly 7,300 addresses as of Aug. 5. The firm flagged a suspected fourth attack wave that could push total losses to 2,055 BTC, or about $130 million.

The Coldcard flaw, traced to a March 2021 firmware change by Canadian manufacturer Coinkite, reduced seed entropy to as low as 40 bits on Mk2 and Mk3 devices versus the intended 128 bits. Block's engineering team estimated the effective search space for Mk4, Mk5, and Coldcard Q devices remained below 73.3 bits.

The Red Team's audit, led by Calle and Rob Hamilton, deployed frontier AI models to scan Bitcoin-adjacent codebases for similar weaknesses. The 85 critical flaws identified across 390 repos suggest the vulnerability extends beyond a single hardware vendor, raising questions about the security posture of the broader Bitcoin open-source ecosystem.

The first attack wave occurred July 30 when an attacker drained 1,082.65 BTC from 1,196 addresses in 41 minutes. Subsequent sweeps followed at roughly 27-hour intervals, with each showing similar patterns of address aggregation. Galaxy said 73 individual victims have contacted the firm directly, and their reports helped researchers identify victim and attacker addresses shared with federal law enforcement, crypto exchanges, and cybersecurity investigators.

Coinkite acknowledged the underlying defect on July 31 and shipped emergency firmware for every affected model. Seeds generated on Mk2 and Mk3 devices without dice rolls carried roughly 40 bits of entropy, and those on the Mk4, Mk5 and Q about 72 bits, against the 128 the devices were meant to produce. Installing the update does not repair a seed already created — users must generate an entirely new seed phrase and transfer their Bitcoin to addresses controlled by the new wallet.

The incident has also prompted debate over whether AI-assisted security testing could have discovered the flaw sooner. Dragonfly managing partner Haseeb Qureshi argued that about $2 of what he called AI hardening could have caught the flaw, citing a test in which the open-source GLM 5.2 model ran for 20 minutes at roughly that cost. However, Tokenomist data lead Tatsapat Saerejittima said some tests were conducted after disclosure and lacked blind testing, documented methodologies and assessments of false positives.

Coinkite's advisory has not changed since Aug. 1 and names no loss figure, no victim count and no number of attackers. "Our investigation is ongoing, and a formal technical review will be released as soon as possible," it says. That review has not appeared.

The Red Team's findings come as the Bitcoin ecosystem grapples with the fallout from the Coldcard incident, which has eroded confidence in hardware wallet security. The 85 critical flaws identified across open-source repos could prompt a broader reassessment of how Bitcoin-adjacent software is audited, with frontier AI models potentially becoming a standard tool for vulnerability discovery.

This article is for informational purposes only and does not constitute investment advice.