The Bitcoin Abstention — Amendment No. 2 | Stellae Liquiditas Research

Sealed as of August 4, 2026 · Stellae Liquiditas LLC · Division 03 · Digital Assets

This amendment supplements and does not replace the firm’s standing position statement, “The Bitcoin Abstention.” Figures stated are current as of the seal date and remain under active revision by third-party forensic analysts.

Purpose

Our standing position rests on three structural liabilities in Bitcoin: a cryptographic exposure it cannot patch on a survivable timeline, a governance model too slow to adapt, and an energy profile antithetical to where the economy is heading. The events of July 30 through August 4, 2026 add a fourth consideration — not a flaw in the Bitcoin protocol itself, but a demonstration of how the asset’s custody culture concentrates risk in ways institutional treasuries must not accept.

We publish this amendment for a second reason. The Coldcard event has been widely mischaracterized in commentary as proof that “cold storage is a myth.” It is not. It is proof of something far more specific and far more useful, and we believe researchers examining custody for any digital asset — not only Bitcoin — deserve a precise account rather than a slogan.

VII. The Coldcard Entropy Failure

What happened

Beginning at approximately 01:10 UTC on July 30, 2026, an attacker swept Bitcoin from more than 1,100 addresses in a single 41-minute window, moving over 1,000 BTC. The sweep was executed roughly 30 hours before the manufacturer published its first security advisory. Galaxy Research, which mapped the initial wave, identified a uniform transaction signature — 30 sat/vB, no change output — and found no comparable pattern in the preceding 30 days.

Estimates have escalated continuously as forensic coverage expanded. Galaxy’s mapping grew to 1,367.05 BTC across 4,585 addresses in subsequent waves; by August 4, Galaxy and Elliptic placed aggregate losses near $130 million, with evidence of at least a dozen distinct actors independently exploiting the same weakness. Galaxy has cautioned that later waves should not be assumed to share an operator with the first, and that the transaction pattern identifies the operator, not the theft — a sweep is indistinguishable on-chain from an owner voluntarily consolidating coins.

These figures are not final. Any party citing this document should treat the loss total as a moving estimate.

The flaw, precisely stated

The vulnerability was not a broken cipher, a compromised supply chain, or a stolen device. It was a build-configuration error that silently disabled the device’s real source of randomness.

Per the security advisory published by Block’s Bitcoin Engineering and Security team, and Coinkite’s own disclosure:

Coldcard’s production configuration sets MICROPY_HW_ENABLE_RNG to zero, because Coinkite supplies its own hardware random-number-generator wrapper rather than using MicroPython’s.
The libngu library checked whether that macro existed rather than whether it was enabled. Because the macro existed — set to zero — the check passed, and the build silently bound to MicroPython’s “Yasmarang” fallback pseudo-random generator instead of the hardware RNG.
That fallback was initialized from the chip’s unique ID and timer registers, and collected no fresh entropy after initialization.
Coinkite estimates effective seed entropy at roughly 40 bits on the Mk3 and about 72 bits on the Mk4, Mk5, and Q — against the 128 bits a 12-word BIP-39 seed is supposed to carry.

The consequence: recovery phrases generated on affected firmware were drawn from a search space small enough to brute-force. An attacker never needed to touch the device. The seed could be reconstructed remotely, from public blockchain data, for any wallet created on vulnerable firmware. Block declined to publish a single practical brute-force figure, setting conditional ceilings and cautioning that these are not equivalent to true cryptographic security of that bit-length.

The defect had been latent in shipped firmware for approximately five years.

Why the remediation matters more than the bug

Coinkite patched the flaw. The patch does not help anyone already affected. A seed generated under weak entropy remains weak forever; updating firmware does not retroactively add randomness to a phrase that already exists. Affected users must generate an entirely new seed on corrected firmware and physically move every coin to new addresses — a migration that costs fees, creates taxable events in many jurisdictions, and depends on each individual user learning of the problem and acting.

This is the same structural pattern documented in our standing position regarding post-quantum migration: Bitcoin’s remediation model is voluntary, per-user, and unbounded in duration. There is no protocol-level key rotation, no administrative recovery, and no mechanism to compel or verify migration. A defect discovered on Monday is still being exploited on Friday because notification and remediation move at the speed of individual human attention.

Contrast: the XRP Ledger supports native, account-level key rotation in its base protocol. A compromised signing key can be replaced without moving funds, changing addresses, or incurring a disposition. This is not a marketing distinction. It is the difference between a recoverable failure and a permanent one.

VIII. What This Event Is Not — A Comparative Case File

Precision in attribution is the entire discipline of security research. The following incidents are frequently grouped together as “cold wallet hacks.” Only one of them is.

Incident	Date	Loss	Actual root cause	True cold-storage failure?
Coldcard / BTC	Jul 30, 2026	~$130M (est., moving)	Firmware entropy defect — weak seed generation at the device level	Yes — device/firmware flaw
Ellipal / XRP	Oct 12, 2025	$3.05M (~1.21M XRP)	User imported seed phrase into the vendor’s mobile app, converting a cold device into a hot wallet	No — user error
BlackWallet / XLM	Jan 13, 2018	~$400K (~700K XLM)	DNS hijack of the hosting provider; injected code auto-forwarded balances over 20 XLM	No — web wallet, never cold
Drift Protocol / SOL	Apr 1, 2026	$285M	Operational compromise of a privileged key or signing credential	No — key/personnel compromise
Upbit / SOL	Nov 2025	~$36M	Exchange hot wallet intrusion	No — hot wallet, custodial
U.S. v. Yaroch (multi-asset)	Charged Aug 2026; conduct alleged early 2025 – Jul 2026	~$1M alleged	Cleared insider located and memorized seed phrases held in government systems	No — authorized insider at the custodian

Five observations follow, and they are the substance of this amendment.

First — only the Coldcard event is a genuine hardware failure. In the Ellipal case, the victim purchased directly from the manufacturer and never disclosed the recovery phrase; the device was not defective and the supply chain was not compromised. The single fatal step was importing the seed into an internet-connected companion app. That is a cold wallet that stopped being cold. The BlackWallet incident involved no hardware at all — attackers hijacked DNS at the hosting provider and modified served code to sweep balances. Calling either a “cold storage breach” obscures the actual lesson.

Second — no confirmed true cold-storage hardware or firmware breach exists for XRP, XLM, or XDC. We state this as a factual observation about the public incident record as of the seal date, not as a claim of superior security. Absence of a documented failure is not proof of invulnerability, and any researcher who treats it as such has learned the wrong lesson from this document. What it does establish is that the frequent rhetorical equivalence — “every chain has cold wallet hacks” — is not supported by the record.

Third — the two largest losses in this table share a root cause, and it is not cryptography. The Coldcard event and the Drift Protocol event are the same failure viewed from opposite ends. Coldcard was a machine-generated key failure: entropy the user could not inspect, produced by a device they were told to trust. Drift was a human-operational key failure: forensic analysis indicates the attacker obtained a privileged signing credential rather than exploiting a flaw in Drift’s program logic, following months of social engineering in which operatives posing as business contacts met employees in person. The funds were swapped through Solana DEX aggregators into SOL and stablecoins, then bridged out toward Bitcoin — the cross-chain exit that analysts identify as a Lazarus Group signature, since Solana-native attackers typically dump on Solana. Attribution to the DPRK-linked Lazarus Group rests on multiple converging axes: laundering pattern, infrastructure overlap, recipient-cluster reuse, and behavioral cadence.

Neither incident was a protocol failure. Both were key-management failures. One at the point of generation, one at the point of authorization. This is where treasury assets are actually lost.

Fourth — the third failure mode is the custodian’s own insider, and it is the one the industry least wants to discuss. On August 1, 2026, a criminal complaint was filed in the Eastern District of Virginia against Patrick Steven Yaroch, a former FBI supervisory special agent assigned to the Counterintelligence and Espionage Division and holding a Top Secret clearance. Prosecutors allege that between early 2025 and July 2026 he transferred roughly $1 million in digital assets from wallets tied to an adversarial nation — accounts he was assigned to investigate — into personal accounts, and charged him with interstate transportation of stolen goods and receipt of stolen goods. Investigators recovered approximately $925,426 into government-controlled wallets; a further sum remained at an exchange. He is alleged to have self-reported before charges were filed. These are allegations. Yaroch is presumed innocent unless proven guilty.

Two details in the affidavit deserve the attention of every treasury operator, and they have nothing to do with the defendant’s motive:

The keys were not exfiltrated. They were memorized. The affidavit describes an insider locating recovery seed phrases within agency systems and committing them to memory. No malware, no exploit, no data-loss-prevention alert to trigger. A BIP-39 mnemonic is designed to be human-memorable — that is the entire point of the encoding — and a secret that one human can memorize is a secret one human can walk out with. Every control premised on monitoring data movement is blind to this.
The custodian was the United States government. Not a startup, not an unaudited offshore exchange. An organization with the most rigorous personnel vetting apparatus in the country, holding assets under a Top Secret clearance regime, failed at insider threat. If clearance-level vetting is not a sufficient control, no vendor’s background-check policy is either.

Fifth — the adversary is a state. Blockchain analytics firms attribute the substantial majority of 2026 crypto theft volume to DPRK-linked operations. A treasury that models its threat surface against opportunistic hackers has mismodeled the problem. The relevant adversary conducts multi-month human intelligence operations against named employees, and it does not need a cryptographic breakthrough when a signing credential will do.

IX. The Learning Moment — Custody Doctrine for Researchers

We publish the following because the commentary around this event has been unhelpful, and because we want the reasoning behind our own custody posture on the record.

First, define the terms correctly

The industry’s “hot versus cold” binary is too coarse to be useful. A more accurate spectrum:

Hot wallet — signing keys reside on an internet-connected device. Exchange accounts, browser extensions, mobile apps. Appropriate only for operational float you can afford to lose.
Warm wallet — keys held offline but routinely connected for signing. Most “cold” hardware wallets in daily practice are warm. This is where the Ellipal victim actually sat.
Cold storage — keys generated and permanently retained on a device that never touches an internet-connected system; signing occurs air-gapped via QR or SD card.
Qualified custody — assets held by a regulated custodian under fiduciary or trust obligations, with segregation of client assets, audited controls, insurance, and institutional key ceremony. In the United States this implicates state trust charters or a national trust charter under OCC supervision, and the SEC’s qualified-custodian framework for regulated advisers.

A necessary correction to the conventional advice. Qualified custody is ordinarily presented as the terminal answer to self-custody risk. It is not. It relocates the risk — from your operational discipline to the custodian’s internal governance — and the Yaroch complaint is the argument for why that relocation must be evaluated rather than assumed. Qualified custody is only superior where the custodian’s architecture makes single-human key reconstruction impossible, not merely prohibited. The correct question to a custodian is not “how do you vet your people?” but “how many of your employees, colluding, would it take to move my assets — and can any one of them ever see a complete key?” If the honest answer is one, the charter and the insurance are decoration.

The critical insight from the Ellipal case: a device’s marketing category is not its operational category. Cold is a property of how the key is handled, not of what the box is called. The moment a seed is typed into anything networked, the device’s classification is irrelevant.

Do

Verify entropy independently. Where the device supports it, supplement or replace vendor-generated randomness with user-supplied entropy (dice rolls, coin flips) and verify the resulting seed derives the expected addresses on independent software. The Coldcard event is the precise argument for this.
Assume the seed is the asset. Every control decision should be evaluated by one question: what expands the number of places this phrase exists, or the number of systems that have ever seen it?
Use multi-signature or threshold schemes for material balances, with keys generated on devices from different manufacturers and stored in different physical locations. Coldcard’s failure was correlated across an entire product line; vendor diversity is what breaks that correlation.
Prefer chains with native key rotation. The ability to replace a compromised key without moving funds is a first-order security property, not a convenience feature.
Subscribe directly to manufacturer security advisories. In this event, the attack preceded public disclosure by roughly 30 hours. Detection latency is a survivability variable.
For institutional balances, use a qualified custodian — and audit their insider-threat architecture, not just their credentials. Evaluate on documented criteria: regulatory status and charter, SOC 2 Type II attestation, proof-of-reserves methodology, insurance limits and named exclusions, HSM and key-ceremony design, and whether client assets are legally segregated from the custodian’s estate in insolvency. Then add the questions the Yaroch matter makes unavoidable: is any single employee ever in possession of a complete key or mnemonic? Is the signing quorum M-of-N across separated individuals? Are seed phrases generated and held such that no human ever reads one in full? Is privileged access logged, dual-controlled, and reviewed by someone outside the custody function?
Prefer architectures where the secret is never human-readable. MPC and threshold signing, or multi-signature across independently held devices, structurally defeat memorization. A key that exists only as distributed shares cannot be committed to memory by one person, because no one person ever sees it.
Model the insider and social-engineering vector explicitly. Drift was not breached through code. Treat “who can authorize a transaction, and who has met them” as a security control.

Do not

Do not import a recovery phrase into any companion application, ever — including one published by the device’s own manufacturer. This single action is the most common root cause of losses labeled “cold wallet hacks.”
Do not assume a firmware patch remediates an already-generated seed. It does not. Migration to a new seed is mandatory after any entropy defect.
Do not equate open-source firmware with verified firmware. Coldcard is open source. The defect survived five years in public view. Security derives from architecture, testing, and independent verification — not from the license.
Do not photograph, cloud-store, type, or transmit a seed phrase. Not in a password manager, not in a note, not in an encrypted archive you sync.
Do not concentrate an entire treasury behind a single vendor, a single device model, or a single firmware lineage.
Do not treat exchange custody as cold storage. The Upbit incident was a hot wallet at a regulated exchange. Custodial convenience is not custodial security.
Do not treat personnel vetting as a security control. It is a screening measure, and screening filters intent at a point in time — it does not constrain capability afterward. The Yaroch complaint describes a Top Secret clearance holder inside the FBI. Architecture constrains; vetting only hopes.
Do not permit any single human — yours or a vendor’s — to be capable of reconstructing a key. If one person can, then your custody model’s real security parameter is that person’s circumstances, and you do not control those.
Do not assume monitoring catches key theft. A memorized mnemonic generates no network traffic, no file transfer, and no alert. Detection controls are largely blind to the highest-consequence insider action.
Do not accept absence of a documented breach as evidence of security — for any asset, including those we hold.

X. Bearing on the Firm’s Position

Stellae Liquiditas holds no Bitcoin, and this event does not constitute our reason for that. We would not have held it had the Coldcard defect never existed. But the event sharpens two points already central to our reasoning.

It demonstrates, concretely, what a network without protocol-level remediation looks like under stress: a five-year-old defect, disclosed 30 hours after exploitation began, remediable only by every affected individual independently discovering the problem and manually migrating. That is precisely the failure mode we described in our standing position regarding post-quantum migration — and it has now been observed rather than predicted.

And it refines our custody doctrine rather than simply confirming it. Stellae Liquiditas holds assets on its own balance sheet as principal. We favor regulated qualified custody with institutional key ceremony over self-custody hardware for core treasury positions, multi-vendor key diversity where self-custody is used, and settlement networks whose base protocol permits key rotation. We do not hold treasury assets on exchange hot wallets beyond operational necessity.

The Yaroch matter obliges us to state the limit of that preference plainly. Qualified custody is our position, but it is not a transfer of responsibility — it is a change in what we are obligated to audit. A custodian earns a treasury position by demonstrating that no single human in its organization can reconstruct a key, not by presenting a charter. We regard insider-threat architecture as a first-order diligence item on every custodial relationship this firm enters, and we will document it as such in our onboarding record for each.

Three events, three points of failure, one category. July 30: the key was generated badly. April 1: the key was authorized by the wrong person. And per the complaint filed August 1: the key was simply remembered by someone permitted to see it. No cipher was broken in any of them.

The lesson of July 30 is not that cold storage failed. It is that entropy is a supply chain, and most holders never audit it — and the lesson of August 1 is that the supply chain does not end at the device. It ends at the last human with permission.

Sources

Consistent with the firm’s Public Content Sourcing Standard, the following are first-party institutional disclosures, forensic-firm analyses, and primary on-chain research. Items sourced only to trade press are flagged as such.

Block, Inc. — Bitcoin Engineering and Security team, security advisory on Coldcard entropy defect (MICROPY_HW_ENABLE_RNG / libngu macro check / MicroPython Yasmarang fallback), July–August 2026.
Coinkite, manufacturer security advisory and firmware remediation notice; stated effective entropy estimates (~40 bits Mk3; ~72 bits Mk4/Mk5/Q), July 31–August 2026.
Galaxy Research, on-chain mapping of the July 30, 2026 sweep: 1,196-address initial wave; subsequent waves to 1,367.05 BTC across 4,585 addresses; 30 sat/vB no-change transaction signature; operator-attribution caveat.
Elliptic (Tom Robinson, co-founder and chief scientist), corroboration of ~$130M aggregate loss estimate, August 4, 2026.
Coinspect, “Ill Bloom” weak-PRNG research (separate, earlier defect affecting legacy software wallets), July 2026.
BIP-39, mnemonic seed specification (128-bit entropy standard for 12-word seeds), Bitcoin Improvement Proposals repository.
ZachXBT, primary on-chain forensic tracing — Ellipal/XRP incident (1.21M XRP; laundering via Bridgers cross-chain bridge into Tron across 120+ transactions), October 2025; and DPRK-linked Solana laundering analyses, 2025–2026.
TRM Labs, DPRK theft attribution and aggregate 2025–2026 crypto-theft volume estimates.
LayerZero, first-party statement attributing the April 2026 cross-chain exploit to a sophisticated state actor.
U.S. Department of Justice, indictment and forfeiture actions concerning DPRK IT-worker infiltration of a Solana trading-bot operator (2026).
Criminal complaint and supporting affidavit, United States v. Patrick Steven Yaroch, U.S. District Court for the Eastern District of Virginia (Alexandria Division), filed August 1, 2026 — charges of interstate transportation of stolen goods, securities, and monies, and receipt of stolen goods, securities, and monies. Referenced for the described method of key compromise only. All statements are allegations; the defendant is presumed innocent unless proven guilty.
BIP-39, mnemonic encoding specification — human-memorable seed representation (design property relevant to the insider-threat vector).
XRP Ledger documentation, account key rotation (regular key / master key) and amendment process.
Flagged — secondary sourcing: the January 2018 BlackWallet (XLM) incident is documented principally in trade press (Cointelegraph; Bitcoin Magazine) and community disclosure; no first-party post-mortem was published. Cited here as illustrative, with sourcing tier noted.

Stellae Liquiditas LLC · Principal-only digital asset treasury · Stellae Group Division 03. Sealed August 4, 2026. This statement reflects the firm’s own treasury conviction and custody doctrine. It is not investment advice, and it is not a security recommendation for any individual’s custody arrangement. Loss figures for the July 30, 2026 event remain under active revision by third-party analysts. References to United States v. Yaroch describe allegations in a pending criminal matter; the defendant is presumed innocent unless proven guilty, and nothing herein should be read as a finding of fact as to any individual.