Cybersecurity in the AI era - Part one

The Coldcard case: what happened

In the seventh episode of Quello che i Bitcoiner non dicono, the Atlas21 podcast, Luca Giuliani – developer at Bitbox – walks through the details of the incident that shook the self-custody world in recent weeks. A vulnerability in the Coldcard firmware, released from March 2021 onward, generated seed phrases with drastically insufficient entropy, opening the door to potential brute-force attacks. According to data available at the time of recording, approximately 1,815 bitcoin were taken from 5,200 distinct addresses, then consolidated into a small number of addresses.

The technical problem has its roots in a business decision. Coldcard, following a dispute with Foundation Devices – which had cloned its code, just as Coldcard had previously done with Trezor – chose to change its firmware licence, making it “source available” rather than open source. That transition forced developers to abandon a well-tested cryptographic library and adopt a less battle-tested alternative. During the migration, an error in the handling of one variable meant the random number generator was not initialised correctly, reducing the effective entropy of the generated seed to as little as 40 bits instead of the intended 128-256.

Why entropy is everything

Giuliani explains the concept of a seed phrase clearly for anyone approaching the topic for the first time. A seed phrase is, in essence, a very large number chosen at random and encoded as a list of 12 or 24 English words. The security of the entire system rests on the vastness of the search space: finding that number by guessing would be like locating a specific atom in the entire universe.

If instead you make some mistakes and do not choose your number with the same degree of randomness, the search space shrinks and it consequently becomes easier to guess.

In the Coldcard case, anyone who knows the bounds of the error can narrow the search to a far smaller space, making brute force economically viable. The only Coldcard users who are safe, Giuliani notes, are those who generated the seed by physically rolling dice – a feature the device supports – or those who added a complex passphrase, which effectively produces a completely different wallet even from a weak seed.

Open source is not enough: the incentive problem

One of the more telling observations in the interview concerns the security culture of the open-source world. The public availability of code, on its own, does not guarantee that anyone will actually read it with a critical eye.

The lesson we take from this affair is that open source is not enough. Over time this created a climate in which nobody was particularly motivated to review Coldcard, because there was no incentive to do so.

Giuliani notes that Coldcard had no bug bounty programme open to external researchers, and that its CEO repeatedly responded with hostility toward anyone who reported vulnerabilities, accusing them of trying to spread FUD. The result was an ecosystem in which nobody had an economic or reputational reason to invest time auditing the firmware – the opposite of what happens in projects where code is freely reusable and where responsible disclosures are rewarded.

How Bitbox approaches entropy generation

Unlike what emerged from the Coldcard case, Bitbox uses what Giuliani calls the defence-in-depth model, inspired by the image of Swiss cheese slices: each layer has its own holes, but when multiple layers are stacked the holes never all line up.

The Bitbox02 combines five distinct entropy sources: the microcontroller’s random number generator, the secure chip’s random number generator, a random value stored in memory during device production that is unique to each unit, entropy supplied by the host device at the moment of use, and the PIN chosen by the user during setup. The combination logic means the resulting entropy depends on the strongest source: even if one of the five were to fail, the remaining ones keep the output at the correct level.

Security in the AI era: a race already under way

The broader context is one of rapid transformation in offensive tools. According to what emerges in the interview, the Coldcard vulnerability was identified using an uncensored large language model, and in parallel a group of researchers funded by OpenSats is using the same tools to find vulnerabilities in open-source Bitcoin projects and report them responsibly to development teams.

Giuliani observes that this is already shifting the relationship between custody and risk. The threshold of bitcoin it is reasonable to keep in a software wallet – a hot wallet connected to the internet – has fallen compared with just two years ago, because finding and exploiting vulnerabilities in code has become cheaper. The guiding principle remains the same: a system is secure when attacking it costs more than it is worth. The costs, however, are changing, and with them the trade-offs to be weighed.

On multi-vendor multisig – multiple keys generated by hardware wallets from different manufacturers – Giuliani agrees it adds a real layer of protection, but warns that the operational complexity carries a concrete cost in terms of human error: forgetting a passphrase or mishandling the backup of one device can lead to loss of funds just as surely as an external attack. The right choice remains tightly dependent on the amount to be protected and the technical ability of whoever manages it.