8/8
AI should detect and explain risk not get unilateral control over funds. Audited policies enforce safeguards, owners keep authority.
Measure detection latency, false positives and losses prevented.
That’s how I would push cold wallet security forward.
1/8
The Ledger/CryptoBilis case has me thinking why does cold wallet security largely stop at protecting the key?
Reports say $92.9M drained from 311 wallets. If someone gets the keys what stops the funds moving?
We need another layer.
7/8
Now the hard part a warning doesn’t stop a stolen seed. Standard single key accounts can’t be remotely frozen.
For prevention, use independent 2 of 3 multisig or enforceable smart account policies: spending limits, delays and extra approvals.
Set up in advance.
6/8
The wallet owner gets a push alert with the exact transaction and next steps. Ledger’s security team gets linked addresses, evidence, severity and affected device cohorts.
That’s actionable intelligence, not a generic “AI detected something” notification.
5/8
Example: 8 enrolled wallets associated with one reseller cohort approve the same unknown spender within seconds.
Not proof of a hardware breach but plenty of reason to trigger a high risk alert and investigate the wider cohort.
4/8
Then stream on chain activity unusual approvals, new spenders, repeat test transfers and multiple wallets moving to the same destination.
Rules catch known signals AI correlates weaker signals across chains and flags incidents for review.
3/8
Maybe offer opt in monitoring inside Ledger Wallet. Register public addresses and privately associate them with verified device/reseller details where available. Never collect seeds.
That gives the security system context the blockchain doesn’t have.
2/8
Ledger already has Transaction Check to flag risky transactions before users sign good.
But if a stolen seed is used outside Ledger Wallet that check never gets a say.
The next step has to go beyond the signing screen.
@ashoKumar89 Knowing what should be built in the first place. Code, deploys and monitoring can all be automated. Context judgement trade offs and understanding why the biz actually needs something are much harder. The developer might write less code but own a lot more of the thinking.
@adxtyahq In a better place than people think. AI or SI will write more of the code sure. But someone has to understand the system make the trade offs spot the bad assumptions, secure it scale it and deal with prod when it gets weird. writing code will be the cheap part Judgement won’t be
The biggest AI companies of the next decade might not be the ones building the smartest models. They might be the ones figuring out how to safely let insanely smart models actually run businesses. Anyway, back to the rabbit hole. 🐇