What was disclosed before?

The page describes itself as a record of "public security research, coordinated disclosures, professional reviews, internal findings, and security advisories affecting COLDCARD", covering 2019 to 2026. Entries are tagged Coordinated, Credited fix, Vendor/internal, Cross-wallet, Overlap, Incident/advisory or AI, and each carries a product scope, a public status and a named reporter where one exists.

Three caveats are the vendor's own, stated on the page, and they are worth preserving because they constrain how the document can fairly be used:

• It presents itself as "a chronology, not a count of independent vulnerabilities or a product score".
• Its "no public evidence" label "means none was found in the cited record; it does not prove that something never happened privately".
• It says raw event counts "should not be used to compare hardware wallets", because older, open and heavily researched products accumulate more public findings.

That last point cuts against a use the document might otherwise invite, and it is correct as far as it goes. The counts still do work for Coinkite, though: a page that opens on 23 events and 12 coordinated disclosures presents a record of engagement with researchers, which is the contested question this history sits inside. V3

The chronology's entry for 1 to 14 March 2022 records a paid private review before the first public Mk4 release, attributed to "Lazy Ninja". It says the review examined SE2-backed PIN derivation, PIN-attempt rate limiting, MCU firewall coverage and random-number generation, and that during the review the runtime random-number generator was seeded with authenticated entropy from both secure elements.

The entry then records something the vendor was not obliged to publish. Further suggestions, including additional bootloader RNG conditioning and reseed hooks, were logged as defence in depth, and in the page's own words "the public firmware history does not show that all were adopted". R4

Two limits keep this in proportion. The review was Mk4 pre-release work in 2022, while the defect at issue here entered through the March 2021 change that routed seed generation through the affected library, and its effect differed by device class, so a recommendation adopted or declined for the Mk4 bootloader does not straightforwardly map onto it. And the entry rests on correspondence this archive cannot inspect, so the scope, the wording of the recommendations and the reason any were not adopted are Coinkite's account of a private document. U5

The chronology records an enterprise AI review of the firmware running from 1 May to 26 June 2026. It reports 85 candidate findings, 8 high, 37 medium, 20 low and 20 informational, each manually reviewed by Coinkite. Most of the high group is described as false positives, intended behaviour, documented trade-offs or unreachable paths, with two exceptions that received fixes. The review is recorded as closed on 26 June, with all items fixed, rejected, accepted as design choices or tracked elsewhere. R6

The theft began on 30 July, about five weeks after that review closed. The chronology does not claim the review covered the seed-generation path, and it lists the flows it did reach: CCC and SSSP, signing, Seed XOR, Key Teleport, PSBT handling, QR, USB, NFC and the WIF Store. On the page's own account, then, a substantial AI-assisted review of this firmware ran to completion weeks before the incident without the seed-generation defect appearing in its output. R7 This is the same class of evidence as the pre-incident and post-incident model tests set out on the AI page, and it carries the same limit: without the prompts and scope, a miss is not measurable.

On 7 August the COLDCARD account added a claim about how the defect was eventually found rather than about what its own review had covered: it said it believes it took the latest LLM models to find it. R8 Nothing in that post adds prompts, scope or transcripts to the 26 June entry, so it does not move the grading above.

Two separate entries do the same work. A 2021 vendor note records that firmware 4.0.0 shipped with a USB serial REPL enabled, and states that 4.0.0 was built, signed and tested internally but never released publicly, making 4.0.1 the first public 4.x binary. The 2026 incident entry then states that Block traces the RNG implementation to the v4.0.0 source lineage, that Coinkite did not publicly release v4.0.0 as a signed binary, and that "the affected-user range therefore begins at v4.0.1". R9

That boundary is not the one Block publishes, or the one this archive records. Block's engineering disclosure classifies Mk2 and Mk3 v4.0.0 as affected, describing the path as first appearing in "released firmware v4.0.0" on 17 March 2021, and the firmware record carries the same published range. The chronology's own wording keeps the two accounts distinguishable: it grants that Block traces the implementation to the v4.0.0 lineage, and rests the narrower figure on distribution rather than on the contents of the build. The disagreement is therefore about whether that build reached the public. R10 Neither party publishes distribution figures, and this archive holds none, so the difference is recorded on both pages rather than settled on either.

James O'Beirne's 4 August first-person account says he audited the firmware in May 2025, became concerned about the RNG path and the libngu constants, and reported it to the COLDCARD team, where he says it was dismissed. R11

The chronology carries no May 2025 entry. Between the March 2022 review and the September 2025 Delta PIN disclosure it records nothing at all, and no entry anywhere in it describes a randomness or libngu report before the incident. V12 It is unchanged on that point: the page has been polled through 8 August and its text hash has not moved since capture.

O'Beirne's account widened on 7 August. Replying to a Coinkite correction about where the weak generator came from, O'Beirne wrote that COLDCARD did fall back to weak entropy, that the copy in libngu was weaker still because it used hardcoded constants, and that "you were warned about this particular issue by me and at least one other person 4 years before me". That is a claim of a second, earlier report, around 2021, by someone he does not name. He linked the libngu source lines carrying the constants; he attached no report or correspondence for either warning. R13

Coinkite's public position the same day runs the other way. Answering a critic, the COLDCARD account wrote that "the bug lived in public for 5yrs, even third party researchers didn't find it", adding that it believes the latest models were needed to find it. R14 A third prior-warning claim, unrelated to O'Beirne's and made by neither party, is also held: a 2 August thread pointing to a video said to show the BTCRecover maintainer warning about COLDCARD entropy two years before the incident. The video is not captured here, and nothing establishes that any such warning reached Coinkite. U15

Neither account settles the other, and this page does not try to. By the chronology's own stated convention an absence means only that nothing was found in the cited record, and the cited record is one the vendor assembled. By the same token a first-person recollection published during an incident, with no contemporaneous artefact attached, is not a document either. What can be said is narrow and worth saying plainly: one party describes a report and now a second, earlier one by another person, the other party's published history of reports contains neither and says researchers did not find the bug at all, and no artefact has surfaced that would decide it. U16

Before the chronology existed, Vlad Costea published a thread arguing that Coinkite's handling of past disclosures showed a pattern, including a 2019 disclosure he says was rewarded with merchandise and a later researcher who he says disclosed to other vendors but not to Coinkite. R17

The two documents overlap on events and emphasise different parts of them. Coinkite's 2019 entries are tagged Coordinated and name their reporters. Costea describes how he says reports were received, a detail the chronology's coordination tag does not address. For the May 2020 factory-reset finding, the chronology lists the mitigation as "disputed", records no bootloader retrofit, and says the linked vendor response is no longer available. V18 Reception is not something either document can settle. Both are held.

The reception question resurfaced on 7 August, with the vendor invoking the standard itself. Answering a critic, the COLDCARD account wrote: "As promised we are doing all the investigation we can to get to the bottom of it. We are devastated. The hacker could have done the right thing and responsibly disclosed." R19 The line was quoted back at the company repeatedly through the day, and more than one reply turned it into the reception question this section is about, one asking whether roughly five separate researchers had disclosed to Coinkite and been mocked for it. No captured artefact enumerates pre-incident disclosures of this defect, so that number is a characterisation and not a count this archive can check. R20