Bitcoin Brief

Sup, freaks.

Let's get into it.

LEAD STORY

The COLDCARD Attackers Left More Than a Blockchain Trail

We have spent the last week covering the COLDCARD disaster aggressively. We traced the failure, the theft waves, the victim reports and the race to secure affected funds. That work matters, but a reader sent me a fair criticism yesterday: the stolen bitcoin is only the surface of this story.

A large part of the story is the industrialization of cyber offense.

The technical failure is now clear. A firmware bug introduced in March 2021 sent affected seed-generation paths through MicroPython's deterministic Yasmarang fallback instead of the intended hardware random-number generator. Coinkite's technical report and Block's independent analysis show that Mk2 and Mk3 firmware added no cryptographic reseed, while later affected devices reduced the secure-element contribution to a 32-bit reseed. That left a sharply constrained universe of possible seeds.

Bitcoin's cryptography did not break. The keys were born weak, and somebody built the machinery to find them.

I think there is a strong probability that AI helped the attackers discover the vulnerable path. I think there is an even higher probability that AI was used to help wage the attack. That is my judgment, not a completed public attribution. Nobody has independently identified the attackers' model, compute provider, host, IP address or cracking program.

The likely workflow is not mysterious. Inspect the public code. Reconstruct the weak generator states. Derive the corresponding keys and addresses. Compare those addresses against funded outputs. Automate the sweeps. Frontier models can compress the code-review and scripting stages. GPUs can accelerate parts of the enumeration. The remaining investigative question is where that work happened and which third parties may have records of it.

There are two broad paths.

The attacker may have been sophisticated enough to download an open-weight model, rent or own the hardware and run the entire operation without exposing prompts to a model company. If so, the evidence may sit with a GPU cloud, hosting provider, payment processor, VPN, email provider or code host. Resource-provisioning events, account-creation records, billing instruments, login IPs, API actions and storage metadata could help identify who controlled the workload.

The attacker may also have used a hosted frontier model or an API. If so, the model provider may hold the original prompts used to review the COLDCARD code, the follow-up requests used to narrow the vulnerability and the conversations used to write or debug the attack scripts. The useful questions are concrete. Which account made the requests? When were they made? Which source IPs and payment instruments were attached? What tools or files were uploaded? How long does the provider retain prompts, outputs and account metadata? Has anybody asked it to preserve those records?

We do not know which path the attackers took. We do know that at least one operator appears to have left a major leak somewhere else in the stack.

Block engineering lead Clay Garrett says one operator used a paid account at an unnamed blockchain-services provider. According to Garrett, the provider's logs matched the number, timing and sequence of the operator's address requests with extraordinary specificity. Garrett said Block was sharing the information with the appropriate authorities.

That operator understood enough about Bitcoin to steal it, but not enough about operational security to hide every dependency. A paid service account may carry an email address, payment record, login history, API key, request body, IP address, device fingerprint or support conversation. Somebody needs to preserve the complete account record before ordinary retention systems erase it.

The blockchain provides the other half of the map.

Galaxy Research's latest formal high-confidence estimate tracks 1,596 BTC from roughly 7,300 addresses across three major waves and smaller owner-confirmed incidents. Those are addresses, not 7,300 victims. Alex Thorn later said the transaction patterns point to at least 15 attackers or operating footprints. That is a clustering estimate, not 15 identified people.

At a fresh Thursday-morning chain snapshot, four publicly named Wave 1 addresses held approximately 1,082.57 BTC in incident-attributed principal and residual balances. The coins are being watched. They are not frozen, recovered or controlled by investigators.

Other operators have already touched services that may know more. MistTrack traced a Wave-4-labeled path that moved 64.90373154 BTC into what it identified as a Wasabi deposit before the deposit was spent in a 382-output CoinJoin-like transaction. Alex Thorn reported that, in a separate theft involving nearly 30 BTC, a 17 BTC tranche moved through THORChain into ETH and a downstream 229.72497255 ETH deposit reached Duel.com. According to Thorn, Duel identified the depositor through its customer records, but the funds had already left.

The chain shows the getaway route. The service accounts may reveal the people.

The people whose bitcoin was taken are the victims of this attack. The goal now should be to help them establish ownership, file reports, preserve devices and transaction records, and give investigators the evidence needed to follow every account the attackers touched.

This is also much bigger than COLDCARD.

Boltz reported months of automated, AI-assisted probing and said attackers began iterating faster than its small team could patch weaknesses. Boltz paused swaps while it hardened the system. ZEUS separately took infrastructure offline after a cybersecurity incident, said the attack was mitigated and said customer funds were neither lost nor at risk. The Bitcoin Red Team has reported thousands of model-generated candidates across hundreds of projects, including many tagged high or critical. Those candidates are not all confirmed vulnerabilities, and the incidents are not all connected. Together, they show that Bitcoin infrastructure is already operating under machine-speed pressure.

The same pressure is coming for every industry.

CBS News reported that cyberattacks had affected water systems in at least 12 states and may be linked to Iran-backed hackers, citing sources familiar with the investigations. The public evidence does not establish that those attackers used AI. I would be shocked if serious offensive operators were not bringing frontier models into reconnaissance, code review, exploit development and automation. At this point, refusing to use the best available models would be a self-imposed disadvantage for an attacker.

I believe we crossed a critical tipping point in July. Kimi K3 arrived as a frontier-capable model, and Moonshot released its model weights and technical report on July 27. That means sophisticated offensive capability is no longer confined to a handful of American labs or mediated by their safety policies. A capable operator can download the model, run it privately and point it at public code without asking permission.

The Bitcoin Red Team is demonstrating the defensive side of the same reality. Calle says the team is spending roughly $10,000 a day on open-weight models including Kimi K3 and Qwen 3.8, while US frontier models did not produce a single vulnerability during the campaign. OpenSats has launched Code RED and a dedicated Red Team Fund to support the work. The raw findings still require human reproduction, triage and coordinated disclosure, but the direction is obvious. Security review is becoming an industrial process too.

That is the defensive posture every serious company now needs. Assume that your public code, exposed services, employee workflows and vendor dependencies are being inspected by tireless agents. Run the same class of models against your own systems. Build a human team that can reproduce findings, prioritize real risk, coordinate fixes and preserve evidence. Know which vendors hold which logs before an incident. Establish the legal and operational relationships needed to move quickly when the first alert arrives.

It is nice to see the Bitcoin immune system reacting so quickly despite the tragedy of the COLDCARD failure and thefts. Researchers are tracing funds. Developers are scanning repositories. Maintainers are patching. Funding is moving toward defense. Other industries should take note and do the same.

The immediate work remains unfinished. Map every model provider, GPU host, blockchain service, exchange, casino, swap service, VPN, payment processor and email account the thieves may have touched. Preserve the records. Identify the operators. Give the victims the best possible chance of getting clarity and, hopefully, getting their bitcoin back.

The era of machine-speed cyber offense is here. Defense has to catch up now.

SIGNAL

BIP-110 Is Entering Mandatory Signaling With Almost No Miner Support

BIP-110 is approaching its mandatory-signaling window at block 961,632 with miner support at roughly 2.5%. The proposal is a temporary one-year soft fork that restricts several methods of embedding arbitrary data in Bitcoin transactions. It caps new non-OP_RETURN scriptPubKeys at 34 bytes, retains an 83-byte limit for OP_RETURN and limits several payload and witness items to 256 bytes. Existing UTXOs are grandfathered.

This is not an established ecosystem-wide consensus upgrade. Start9 warns that enforcing nodes could stall or follow a competing chain if hash power does not adopt the rules. That is a plausible failure mode, not a prediction. But with 41 of 1,668 tracked blocks signaling, anyone running enforcing software needs to understand what happens at the boundary.

Wavelength Packages Ark, Lightning and Bitcoin Into One Self-Custodial SDK

Lightning Labs released Wavelength, an early developer system that combines an Ark client, a Lightning swap engine and an on-chain bitcoin wallet behind one daemon. It supports web, React Native, native iOS and Android, REST and gRPC APIs, and a command-line interface.

Users hold their keys locally. Lightning payments are handled through atomic swaps from Ark balances, which means users do not have to manage Lightning channels. The MIT-licensed repository also includes an MCP server that exposes wallet functions to AI agents as typed tool calls. This is not proof of production adoption. The public quick start defaults to signet, and routine Ark use still depends on operator and swap infrastructure. But the direction is right: make self-custodial bitcoin programmable without forcing every application team to become a wallet-infrastructure company.

Bitcoin Is Currently the Seventh-Largest Base Money in Mežinskis's Ranking

Matthew Mežinskis published his 32nd global base-money update Thursday. His March 31 ranking puts gold first at $28.4 trillion, followed by the Chinese yuan at $5.9 trillion, the US dollar at $5.4 trillion, the euro at $4.8 trillion, the Japanese yen at $3.7 trillion, silver at $2.6 trillion and bitcoin at $1.4 trillion.

Bitcoin sits ahead of the British pound's approximately $990 billion monetary base and every other fiat monetary base below sterling in Matthew's March 31 table.

The denominator matters. M1, M2 and M3 are broader aggregates that include currency, bank deposits and other liquid claims; they are not the same measure as central-bank base money. Matthew estimates $146.39 trillion of global broad money against $26.41 trillion of fiat base money at year-end 2025, a ratio of 5.54-to-1.

His 95-member, USD-size-weighted compilation reports an 11.2% all-time CAGR from December 31, 1969 through March 31, 2026, even though the trailing twelve-month reading was -2.1% in US-dollar terms. Bitcoin is currently seventh without a committee expanding its supply. A brutal week does not change where it sits in the monetary hierarchy.

The BLADE Act Targets Industrialized Theft of AI Capabilities

Senators Bill Hagerty, Tim Scott and Andy Kim have introduced the bipartisan Blocking Large-scale Adversarial Distillation Efforts Act. The sponsors say foreign adversaries are using large-scale querying and adversarial distillation to reproduce capabilities developed by American AI companies.

The proposal aims to expose and punish the entities behind those extraction campaigns. That connects directly to the defensive problem in today's lead. Trusted model services can become attack surfaces, but they can also hold the request patterns, accounts and payment records that reveal a coordinated campaign. Introduction is not passage, and the initial announcement is heavier on intent than enforcement mechanics. Still, Washington is recognizing that offensive actors do not need to steal model weights if they can industrialize model access itself.

Cook Says Inflation Risk Now Outweighs Labor Risk

Federal Reserve Governor Lisa Cook says the balance of risks has shifted toward inflation. Twelve-month PCE inflation through June was 3.7%, core PCE was 3.3% and unemployment was 4.2%. She also pointed to AI-infrastructure spending as a source of higher prices for semiconductors, high-tech equipment, software and utilities.

Cook supported holding rates steady in July, but she said she is prepared to raise rates if disinflation does not resume. The labor market is stable enough, in her view, to absorb tighter policy if necessary. This is one governor's position, not an FOMC commitment. It is still a warning for liquidity-sensitive markets: the AI capex boom is not automatically disinflationary, and the Fed is not obligated to look through the buildout if it keeps goods and utility prices elevated.

⚡ FREEDOM TECH CORNER

Code RED Turns Open Models Into Bitcoin's Defensive Arsenal

The Bitcoin Red Team is moving from an emergency campaign toward durable defensive infrastructure.

Calle reported Thursday that US frontier models did not produce a single vulnerability during the campaign, while the team was spending roughly $10,000 a day on open-weight Kimi K3 and Qwen 3.8. That is a participant report, not an independent benchmark, but it exposes the problem. Defensive researchers cannot wait for a permission gate while attackers can download open-weight models and run them privately against the same public code.

OpenSats has launched Code RED and a dedicated Red Team Fund to reimburse LLM token costs and support security researchers. That is exactly the response the ecosystem needs. Open models give defenders the same private, tireless review capacity available to attackers. Funding turns emergency scanning into an enduring capability.

The machines can generate findings. Humans still have to reproduce the bug, determine whether it is exploitable, coordinate disclosure, help maintainers patch it and avoid publishing a roadmap for criminals. Bitcoin's advantage is not that it avoids attack. Its advantage is that the code is open, the stakes are visible and the immune system can learn in public.

DATA SNAPSHOT

As of August 6, 2026, 12:17 p.m. ET

Sources: Coinbase; mempool.space; Farside Investors; Galaxy Research.

See you tomorrow. This is not investment advice. Do your own research.

YouTube: https://www.youtube.com/@TFTC

podcast: https://www.tftc.io/tag/podcasts/