Over the past month, a series of incidents has once again put the crypto industry on high alert.
First, Coldcard was found to have a critical vulnerability in its random number generation. Trezor and SafePal then disclosed separate incidents that involved potential exposure of users’ personal data.
At first glance, the three incidents appear to have little in common. But viewed over a longer timeline, they point to an increasingly important question:
As AI automates vulnerability discovery, exploit development, and social engineering, how many parts of a crypto wallet could become the next weak link targeted by attackers?
1. How AI Is Turning Hacking From a Craft Into an Industry
Objectively speaking, the three incidents exposed entirely different attack surfaces.
Coldcard’s issue affected private key generation and represented a serious security vulnerability. Trezor’s incident involved a third-party logistics provider, while SafePal’s involved its ordering system and browser extension permissions. The latter two primarily exposed risks stemming from personal data leaks.
Although there is currently no evidence that AI was directly involved in all three incidents, one fact is becoming difficult to ignore: in the AI era, the hacker’s toolbox is undergoing a profound transformation.
Many sophisticated cyberattacks were previously constrained by a very practical resource—human time.
Studying a large codebase, understanding its call paths, and identifying logic flaws hidden for years could require an experienced security researcher to invest enormous amounts of time. Collecting information on a particular user, studying their habits, and crafting a convincing phishing email could require months of preparation to develop a sophisticated social-engineering campaign.
This forced attackers to make a trade-off. They could either automate an attack at scale using relatively crude methods and wait for a small number of victims to take the bait, or meticulously target a high-value individual with an attack that was difficult to replicate at scale.
As AI capabilities rapidly advance, however, that toolbox is receiving a complete overhaul:
- Automated vulnerability discovery: AI can help attackers rapidly analyze smart contracts, client software, and even firmware to uncover zero-day vulnerabilities and logic flaws.
- Scalable social engineering: Phishing emails that once required careful manual preparation can now be generated automatically using leaked identity data. AI can produce highly personalized and persuasive emails, text messages, voice recordings, and even videos. Further reading: Spring Festival Asset Security Guide: Protecting Your Tokens Amid the Holiday Rush
- Intelligent attack execution: From identifying targets to launching coordinated attacks across multiple channels, the cost of executing an entire attack chain has fallen to an unprecedented level.
Capabilities that were once distributed across different specialists—from target selection and vulnerability research to malware generation, social engineering, and attack delivery—are gradually being compressed into a more automated workflow.
This is the truly far-reaching impact of AI on cybersecurity.
AI may not suddenly invent an entirely new form of attack. Instead, it rapidly lowers the cost of existing attacks. Finding a vulnerability becomes cheaper, analyzing a target becomes faster, and generating a thousand different versions of a phishing email becomes far easier than before.
In other words, the fact that many systems were not attacked in the past did not necessarily mean they had no vulnerabilities. Sometimes, those vulnerabilities were simply too difficult to find, the attacks were too expensive to execute, or the potential targets were not worth the effort.
The invisible security boundary created by the assumption that “attackers do not have enough time” is now growing thinner.
From this perspective, the battle over crypto asset security is expanding beyond the relatively narrow contest for private keys. It is becoming an end-to-end struggle spanning code, devices, supply chains, user identities, and everyday interactions.
AI is simply accelerating that shift.
2. A Wallet’s Attack Surface Extends Far Beyond Its Seed Phrase
This is why the recent incidents are particularly revealing when examined together.
Each affected a different stage of the wallet lifecycle, reminding us that wallet security has long since moved beyond the single question of whether a private key has been stolen. Risk can be embedded in every step, from private key generation and hardware devices to logistics and users’ personal information.
Let us break down the three incidents.
Coldcard is the most direct example. Its vulnerability arose before users had even begun using their wallets.
A seed phrase could still appear as 12 or 24 ordinary words. The device could sign transactions and transfer assets normally, leaving the user with little reason to suspect anything was wrong. But if the randomness used to generate that seed phrase was not truly random, the wallet could still be at risk—even if the user had never shared the phrase with anyone.
After all, the advice to “protect your seed phrase” assumes that it was generated securely and unpredictably in the first place.
Trezor and SafePal illustrate a different type of risk.
Unlike Coldcard, their hardware was not compromised, and users’ seed phrases remained intact. What was exposed instead was purchase information, including names, phone numbers, email addresses, and even delivery addresses.
Imagine purchasing a top-of-the-line, tamper-resistant safe. The safe itself remains secure, but the shipping company loses a delivery manifest listing your name, email address, phone number, home address, and the fact that you purchased a hardware wallet designed specifically to store crypto assets.
An attacker would then possess a list of potentially high-value crypto users. They could impersonate wallet support staff and send an “urgent firmware update” notice, create a phishing page tailored to the exact wallet model purchased, call users about an alleged order issue, or link their social media profiles and public identities to onchain addresses.
In other words, being unable to break the cryptography does not leave an attacker without options.
The crypto community has long used an extreme real-world example to illustrate this point: the “$5 wrench attack.” No matter how strong the encryption is, it cannot prevent an attacker from directly targeting the person who owns the assets.
This is not purely a theoretical risk. According to data Chainalysis provided to the Financial Times, at least 46 violent attacks against crypto holders had been recorded by mid-August 2026. Kidnappings accounted for more than half of them, while home invasions represented over one-third.
Looking back at the three wallet incidents, it becomes clear that “wallet security” now involves a very long chain:
It begins with wallet code, randomness, and key generation, then extends to chips, firmware, and devices, followed by official websites, purchasing channels, supply chains, logistics providers, and order databases. Once a user begins using the wallet, it connects to RPC services, DApps, browser extensions, and smart contracts. It must then handle approvals, signatures, customer support, social media, and even AI agents.
If any link becomes the weakest point, attackers may be able to bypass the defenses built throughout the rest of the chain.
3. As Attacks Become Automated, Defense Must Embrace AI
If AI continues advancing at its current pace, the problems being exposed today may be only the beginning.
One of AI’s greatest strengths is its ability to examine a large system continuously for anomalies, recurring patterns, and weak points.
Attackers can deploy agents to scan open-source code around the clock, test websites, APIs, and browser extension permissions at scale, and automatically collect information from social media and public databases before identifying potentially high-value targets.
Even phishing could evolve beyond generic messages such as “Your wallet is about to expire—please enter your seed phrase” into real-time conversations tailored to the victim.
- If an attacker knows that you recently purchased a particular hardware wallet, AI can generate a “firmware security notice” tailored to that exact model.
- If the attacker knows that you recently used a particular DeFi protocol, AI can impersonate the project team and instruct you to migrate your assets to a new protocol vault.
- If the attacker also gathers information from your social media profiles and public posts, AI may even imitate a team member, KOL, or customer support representative you already know and trust.
From this perspective, wallets face an important challenge: when attacks evolve from fixed patterns into systems that can analyze, reason, and adapt, can defense continue to rely on static rules alone?
Traditional wallet security mechanisms still largely resemble a rulebook. If an address has been identified as a phishing address, the wallet displays a warning. If a domain has been blacklisted, access is blocked. If a particular approval pattern is considered risky, an additional alert appears.
These mechanisms remain important. But as attacks become increasingly dynamic, identifying the next threat solely by looking at threats that have already occurred is clearly not enough.
AI can become an important addition to the defensive toolkit. Further reading: When Hackers Scale Up with AI: The Next Level of Web3’s Security Arms Race
In fact, this is not an entirely new idea.
In previous discussions about “AI × Web3 security,” imToken proposed a similar direction: wallet security should move beyond address blacklists, risk labels, and fixed pop-up warnings. With AI, security checks could be integrated earlier and more deeply into the user’s entire transaction process.
Before code reaches production, for example, AI could continuously review dependencies, call paths, and anomalous logic. When a user visits a DApp, the wallet could assess its domain history, front-end behavior, contract addresses, and onchain relationships to detect suspicious activity. Before a signature is submitted, it could simulate the transaction’s actual outcome instead of merely displaying an incomprehensible string of hexadecimal data.
Over time, wallets could even build dynamic security models for individual users.
If an account that normally transfers only a few hundred dollars suddenly attempts to approve a contract deployed just two hours earlier to spend all of its assets, that is an anomaly in itself.
If a user is about to grant unlimited token approval to an address they have never interacted with, the wallet should issue a higher-priority risk warning.
And if an email claims to come from an official wallet team and asks the user to enter their seed phrase, it should be classified as high risk—no matter how convincing the message appears.
The change brought by AI may therefore extend far beyond automatically telling users whether an address is safe. It could give wallets, traditionally passive tools for key management and transaction signing, an active layer of risk assessment.
This also makes another security boundary previously discussed by imToken even more important: AI can help users understand and execute complex operations, but control over assets must not be delegated without limits.
Critical actions—such as large transfers, approvals for new addresses, and interactions with sensitive contracts—should remain subject to least-privilege controls, human confirmation, pre-execution simulation, and clear explanations. AI’s capabilities must stay within explicitly defined permissions.
Most importantly, when something genuinely appears abnormal, the wallet should clearly explain why it is dangerous, what will happen if the transaction is executed, and where the risk lies.
In other words, the value of AI-powered defense lies in transforming wallets from passive signing tools into systems capable of actively understanding transactions, identifying anomalies, and constraining execution.
Final Thoughts
The recent series of wallet security incidents does not mean self-custody has lost its value. Nor does it suggest that users should return complete control of their assets to centralized platforms.
What these incidents remind us is that self-custody has never meant automatic security. It means returning ultimate control over assets to the user.
Protecting that control requires a security system capable of evolving alongside changing threats. Security is not a one-time product deliverable. It is an ongoing and dynamic process that requires users, projects, and wallet providers to work together.
Attackers can use AI to understand code, users, and their environments.
Defenders can do the same.
The long-running “sword and shield” arms race has entered its next stage.