A volunteer group of roughly 20 to 25 Bitcoin developers has been racing to find security flaws in open-source wallet software before attackers armed with AI models find them first. The effort, known informally as the Bitcoin Red Team, formed after AnchorWatch CEO Rob Hamilton began combing through Bitcoin projects in the wake of the Coldcard air-gapped wallet exploit that drained more than $130 million from users starting in July.
Team member Calle, who goes by a pseudonym, said the group doesn't wait to be asked.
“We get a bunch of inbound requests from projects that want to be scanned, but we act proactively, and we've covered almost the entire significant open-source ecosystem by our own sweeps already.”
The team includes both pseudonymous contributors — Calle, Stu, Talip, and thesimplekid — and named developers Ben Carmen, Daniela Brozzoni, James O'Beirne, and Bruno Garcia. As of early August, the group had spent roughly $20,000 on AI services to run its scans.
The Coldcard Exploit That Triggered It
The exploit that spurred the group's formation traced back to a firmware bug dating to 2021, when Coldcard's seed-generation process was quietly rerouted from the device's dedicated hardware random number generator to a software fallback algorithm called Yasmarang — originally intended only for devices with no randomness chip at all. That mistake shrank the effective entropy of some wallet seeds from 128 bits down to roughly 40, making the underlying private keys guessable rather than cryptographically secure.
Attackers began exploiting the flaw in July, draining 594 BTC — about $38 million at the time — from roughly 500 wallets in a single 25-minute window during the first wave. By mid-August, confirmed losses across at least a dozen separate threat actors had climbed past $130 million. Coinkite, the company behind Coldcard, has since released firmware updates that require users to supply their own entropy — through keypresses, dice rolls, or coin flips — combined with hardware randomness when generating new seeds, and has warned that any seed generated under the old method needs to be abandoned and replaced entirely.
Related: Coldcard Overhauls Firmware After $130M Bitcoin Seed Exploit
Why AI Changes the Threat Model
Calle's central warning is about accessibility rather than novelty: the exploit itself required deep technical knowledge, but AI tools now let far less skilled attackers replicate that work.
“Simple exploits can now be completed end to end by someone who doesn't know how to do it without AI.”
He was careful to note that the vulnerabilities the team keeps finding sit in applications and wallet software, not in Bitcoin's core protocol itself — a distinction that matters for how much confidence users should place in Bitcoin generally versus specific products built on top of it.
Notably, Calle said the team leans on Chinese AI models over American ones for its security scans, calling the gap “not even close” because U.S. models tend to impose guardrails that limit their use for offensive cybersecurity testing — even when that testing is defensive in intent. That's created an odd dynamic where the developers racing to protect Bitcoin's software ecosystem are, in effect, using some of the same categories of tools that make attackers more dangerous, just pointed in the opposite direction.
