Bitcoin’s Red Team Leans on Chinese AI: 4,962 Bugs, A Stress Test for Open-Source Bitcoin

Bitcoin’s open-source stack just met a new kind of adversary: permissive, locally runnable Chinese AI. The Bitcoin Red Team, led by pseudonymous developer Calle, has been sweeping wallets, Lightning applications, libraries, and related projects using models like Moonshot AI’s Kimi K3 and Z.ai’s GLM 5.2—supplemented, when allowed, by U.S. systems from OpenAI and Anthropic. The output is uncomfortable but useful: thousands of issues quietly handed to maintainers before public disclosure.

The numbers explain the urgency. In August, the group reported 4,962 findings across 390 projects, including 85 labeled critical and 635 high severity. Calle said developers have confirmed “a ton of real critical and high vulnerabilities.” Their assessment of the landscape was blunt: “We’re experiencing a massive collision between decades of human open source slop against 2 weeks of Kimi K3… Everything is broken, Bitcoin is burning.” That framing sounds dramatic, yet it captures how technical debt behaves when automation finally gets enough context window and reasoning ability to tear through old code.

Why pivot to Chinese models? Access and control. Kimi K3 can be downloaded and run on local hardware, chew through large codebases, and complete long-form tasks with minimal supervision. By contrast, American systems often throttle or block security workflows. Calle’s update—“Red team rugged by OpenAI cyber again… loading up Kimi K3”—reflects a practical trade-off: speed and sovereignty over tooling tend to beat waiting on policy exceptions. Hugging Face faced a similar constraint last month, turning to China’s GLM 5.2 to analyze a breach after OpenAI models hacked its systems and U.S. commercial models refused to parse the attack logs.

Two points matter from here:

- Maintenance is destiny. Calle warned that unmaintained projects should not be trusted. Response times to private reports vary widely, which, in practice, maps to organizational health. Teams that began AI audits months ago are “in a completely different position” than peers, and Lightning software—already complex—looked “more broken than the average.” If your project sits anywhere near payment flows or key management, you likely need a standing AI-powered audit pipeline, not a one-off scan.

- The easy wins are gone. “We’ve basically completed a basic scan of virtually the entirety of Bitcoin open source. The low hanging fruit is done.” That implies the next phase shifts from volume to depth: chain-of-thought code reasoning, cross-repo dependency mapping, and adversarial fuzzing integrated into CI. It also implies heavier cognitive load for maintainers. AI raises the floor on what gets found—and the stress level of those who have to fix it.

There is a strategic undercurrent here: model choice is becoming part of security posture. Locally hosted, less policy-constrained systems empower red teams to move without asking permission, but they also introduce governance questions—data handling, provenance of model weights, and the optics of routing sensitive analysis through foreign stacks. Teams can balance this by keeping scans local, strict-scoping any proprietary inputs, and documenting reproducibility. In my experience, reproducibility and measured disclosure earn faster maintainer engagement than loud severity labels.

Lightning’s pain point is expected. Protocol complexity and asynchronous state channels create more surface area, and legacy code paths persist longer than people admit. Projects that invested early in AI audits seem to be absorbing this shock better. That is the lesson worth internalizing: set up your own AI audit pipeline now—model orchestration, triage heuristics, human-in-the-loop review—so you aren’t recalibrating under fire.

Calle’s “Bitcoin is burning” line unsettles some readers, but the process is normal for maturing ecosystems. Security clarity arrives in waves—painful at first, stabilizing later—especially when new tools expose old assumptions. “Bitcoin is the obvious first target, but the rest of the world will follow shortly.” If this sweep pressures maintainers to retire unmaintained modules, tighten Lightning implementations, and codify AI-first review practices, the codebase gets healthier. Sometimes old things do need to burn so new things can grow on healthier soil.