3 August to 9 August 2026
This week gave me the same reminder from two very different places: the thing with the reassuring label is rarely the whole system.
A hardware wallet can be air-gapped and still protect a key that was weakly created. An open-weight model can carry the same name at two API providers and still behave differently in the work that matters. In both cases, the user sees the familiar surface. The risk sits lower in the stack.
Same model, different work
I have been thinking about open weights mostly as a choice story: more providers, more pricing pressure, more portability. That is still true. But Artificial Analysis’ Endpoint Accuracy Index adds an uncomfortable qualification: the same published weights do not guarantee the same useful output.
The index compares serverless endpoints with a reference deployment on tool calling, hard reasoning and long-context recall. Its early results show why this matters. For gpt-oss-120b, some endpoints scored 22% on the tool-use test against 37% for the reference. In other cases, tight output limits cut off reasoning before a model had finished its work.
That makes a cheap rate card less straightforward than it looks. A provider may be cheap because it is efficient. Or it may be cheap because it produces a shorter, lower-effort answer and leaves the failed work for the next retry.
The practical change for me is small but useful: price is now one column in the model-routing decision, alongside endpoint quality, max output, tool-call reliability and a periodic canary test. “Same model” is no longer enough provenance.
The Coldcard incident: air-gapped is not a provenance check
The sharper example came from Bitcoin custody. A defect in some older Coldcard firmware weakened the randomness used when generating wallet seeds. The device could remain offline, but an attacker could still derive vulnerable keys without ever holding the owner’s wallet.
The first sweep reportedly moved about 594 BTC from roughly 500 wallets in 25 minutes. By 5 August, TRM Labs estimated approximately 1,816 BTC, around $116m at the time, had been taken across more than 5,200 addresses. Those figures were still preliminary, and attribution remains unresolved.
The operational point is not that self-custody is broken. It is that self-custody relocates responsibility. The device protects a seed after it exists. It does not prove that the seed was created with enough entropy, that firmware behaved as expected, or that an old key is safe just because the current firmware is patched.
Coinkite’s own remediation is blunt: updating protects future seed generation. It does not repair an affected existing seed. Affected holders need a new seed and a migration, not merely an update.
What I almost missed
I follow crypto and AI developments every week as part of how I work. Most weeks I am watching models, policy and stablecoin rails closely. Custody and wallet security were in my peripheral vision, until they weren’t.
Coldcard was the reminder. A firmware flaw that weakens how seeds are created is not just “another hack headline.” It changes the controls that still matter after the news cycle moves on: how a key was made, whether an old wallet needs migration, how many independent checks you trust before you move funds.
So I am being more deliberate about those stories, not every exploit, but the ones that change a reusable control: provenance, migration, signer diversity, approvals or incident response.
The parallel is the same one as the models. In AI, do not stop at the model name. In custody, do not stop at the hardware-wallet name. Ask what sits beneath the interface, how it was made, and what would tell you it has changed.
One question I am carrying into next week: which part of the stack am I treating as guaranteed merely because the brand on the box is familiar?