Web3 Wallets in "Turbulent Times": In the AI Era, How Should We Understand the "Spear and Shield" Evolution of Crypto Security?

Over the past month, the crypto community's security nerves have been stretched taut once again.

First, Coldcard was exposed for a severe random number generation vulnerability, quickly followed by Trezor and SafePal disclosing incidents of user data privacy leaks.

At first glance, these three incidents don't seem to share much in common, but if you stretch the timeline a bit further, they collectively point to an increasingly critical issue:

As AI automates vulnerability discovery, exploit development, and social engineering at an accelerating pace, just how many points of weakness remain in a crypto wallet that attackers could target next?

1. With AI, Hacking Transforms from a "Craft" into an "Industry"

Objectively speaking, these three incidents exposed completely different attack surfaces.

Coldcard's problem lies in private key generation—a severe security flaw. Trezor's issue stemmed from a third-party logistics service, while SafePal's was tied to its order system and plugin permissions—both representing risk exposure linked to privacy leaks.

Although there is currently no evidence directly linking all three incidents to AI, it must be acknowledged that in the age of AI, the hacker's "toolbox" is undergoing a dramatic transformation.

In the past, many sophisticated cyber attacks were fundamentally constrained by a very practical limitation—human time.

Analyzing a large codebase, understanding call relationships, and hunting for logic vulnerabilities hidden for years required experienced security researchers to invest significant time; gathering identity information on a specific user, studying their habits, and crafting a sufficiently convincing phishing email could even take months to construct elaborate social engineering scripts.

This meant that past attacks often involved a trade-off: either they were highly automated but crude in approach, casting a wide net hoping a few users would take the bait; or they were meticulously designed for specific high-value targets, but difficult to replicate at scale.

But today, with AI capabilities evolving at breakneck speed, the hacker's toolbox has been completely upgraded:

• Automated vulnerability discovery: AI can assist attackers in rapidly analyzing smart contracts, client-side code, and even firmware, automatically hunting for zero-day exploits and logic flaws.
• Scalable social engineering: Phishing emails that once required careful crafting can now be auto-generated by AI based on leaked user identity data, producing highly customized, persuasive phishing content, SMS, and even voice/video messages (see also: Spring Festival Asset Security Guide: How to Safeguard Your Tokens While Relaxing with Friends and Family?).
• Intelligent attack execution: From target identification to multi-channel concurrent deployment, the cost of the entire attack chain has dropped to an all-time low.

In essence, from target selection and vulnerability research to malicious code generation, social engineering, and attack content delivery, capabilities once scattered across different attackers are being compressed into an increasingly automated workflow.

This is the truly profound impact AI has on cybersecurity.

It may not suddenly create a never-before-seen attack vector, but rather rapidly lowers the cost of existing attack methods—finding a vulnerability is cheaper, analyzing a target is faster, and generating a thousand different versions of a phishing email is far easier than ever before.

In other words, many systems in the past were not attacked not necessarily because they lacked vulnerabilities, but sometimes simply because the vulnerabilities were too hard to find, the attack cost too high, and the return on investment for attackers too low. Now, that invisible security boundary once formed by the assumption that "attackers don't have enough time" is gradually thinning.

From this perspective, the security battleground for crypto assets is also expanding from what was once a relatively concentrated "battle for private keys" into a full-chain tug-of-war encompassing code, devices, supply chains, user identities, and daily interactions.

AI is simply pressing the accelerator even harder.

2. The Real Attack Surface of a Wallet Goes Beyond a Seed Phrase

This is also why viewing the recent spate of incidents together makes them particularly representative.

They sequentially struck different points in the wallet lifecycle, reminding us that the risks facing a wallet have long transcended the single dimension of "whether the private key was stolen," and are instead embedded in every stage—private key generation, hardware devices, logistics supply chains, and even user privacy information.

Let's break this down.

Coldcard is the most typical example—its problem occurred even before the user truly began using the wallet.

The seed phrase still appears as 12 or 24 normal words, the device still signs and transfers normally, and users may hardly notice anything unusual. But if the randomness generating that seed phrase isn't truly random, then even if you never share your seed phrase with anyone, you remain at risk.

Because the very premise of "keeping your seed phrase safe" assumes that the phrase was first generated through a sufficiently secure and unpredictable process.

Then there's Trezor and SafePal.

Unlike Coldcard, their hardware wasn't compromised, and the seed phrases remained intact. However, users' purchase records were leaked—including names, phone numbers, email addresses, and even shipping addresses.

This is akin to buying a top-tier explosion-proof safe. The safe itself wasn't pried open, but the logistics company's shipping manifest was lost, clearly showing your name, email, phone number, home address, and the fact that you purchased a hardware wallet specifically designed to store crypto assets.

What attackers gain is a list of potential high-value crypto users. They can impersonate wallet customer support to send "urgent firmware upgrade" notices, craft phishing pages tailored to the purchased model, call claiming order anomalies, and even go further by correlating users' social media accounts, public identities, and on-chain addresses.

In other words, failing to crack the cryptography doesn't mean the attack has no path forward.

In the real world, there's even an extreme saying that has circulated in the crypto community for years—the "$5 wrench attack": no encryption algorithm, no matter how strong, can solve the problem of attackers simply going directly to the asset holder.

This isn't entirely a theoretical risk. According to data Chainalysis provided to the Financial Times, as of mid-August 2026, at least 46 violent attacks against crypto holders have been recorded this year, with kidnappings accounting for more than half and home invasions exceeding one-third.

So looking back at these three incidents, it becomes clear that what we call "wallet security" today has actually become a very long chain:

From wallet code, random number generation, and key generation, to chips, firmware, and devices, then to official websites, purchase channels, supply chains, logistics, and order databases; after users truly start using it, the chain extends to RPC connections, DApps, browser plugins, and smart contracts, followed by authorizations, signatures, customer support, social media, and even AI agents.

If any single link becomes the weakest point, it could bypass the security defenses established by all other links.

3. When Attacks Become Automated, Defense Must Integrate AI

If AI continues evolving at its current pace, the problems exposed today are likely just the beginning.

Because one of the things AI excels at is continuously scanning a massive system to identify anomalies, recurring patterns, and weak points.

Attackers can deploy agents to continuously scan open-source code, batch-test websites, APIs, and plugin permissions, automatically gather information from social media and public databases, and then filter for potential high-value targets.

Even phishing itself may evolve from the cookie-cutter "your wallet is about to expire, please enter your seed phrase" into real-time conversations that genuinely know who you are:

• If an attacker knows you just purchased a hardware wallet of a specific model, they can generate a "firmware security notice" tailored to that exact model;
• If they know you recently participated in a DeFi protocol, they can impersonate the project team and lure you into migrating to a new protocol vault;
• If they've further obtained your social accounts and public statements, they can even mimic team members, KOLs, or customer service representatives you're familiar with to communicate with you;

From this perspective, a key challenge wallets will face in the future is this: when attacks have evolved from "fixed rules" into systems that analyze, adapt, and change, can defense continue to rely solely on static rules?

After all, wallet security mechanisms have so far remained relatively close to a "rule library": if an address is flagged as a phishing address, a pop-up warning appears; if a domain is blacklisted, access is blocked; if a certain authorization pattern is deemed high-risk, an additional warning layer is added.

These mechanisms remain important, but when faced with increasingly dynamic attacks, relying solely on already-known risks to identify the next one is clearly insufficient.

AI can serve as a crucial complement on the defensive side (see also: As Hackers Leverage AI for Greater "Efficiency," How Does the Web3 "Arms Race" Between Spear and Shield Escalate?). In fact, this is not an entirely new proposition that has suddenly emerged.

Previously, in discussions around "AI × Web3 Security," imToken already proposed a similar direction: the security capabilities of wallets in the future should not remain limited to address blacklists, risk labels, and fixed pop-ups. Instead, AI can be leveraged to push security judgments further upstream into the user's entire transaction process.

For example, before code enters the production environment, AI can continuously audit code dependencies, call paths, and anomalous logic; when a user visits a DApp, it can assess abnormality by combining domain history, front-end behavior, contract addresses, and on-chain association relationships; and before signing, it can simulate the actual outcome of executing the transaction, rather than merely presenting users with an incomprehensible string of hexadecimal data.

Going a step further, wallets could gradually build dynamic security models tailored to each individual user.

An account that has historically only transferred a few hundred dollars suddenly attempting to authorize all its assets to a newly deployed, two-hour-old unfamiliar contract is itself an anomaly signal; an address the user has never interacted with suddenly requesting unlimited-amount approval should warrant higher-priority risk alerts; and an email claiming to be from the wallet's official team, asking the user to enter their seed phrase—no matter how realistic it looks—should be directly flagged as high-risk.

Thus, the change AI brings may not merely be "automatically helping users determine whether an address is safe." It's more like transforming the wallet from a relatively passive key-keeping and signing tool into one that gradually possesses a layer of proactive risk-assessment capability.

This also makes another security boundary imToken previously discussed even more important: AI can help users understand and execute complex operations, but control over assets must not be infinitely ceded as a result. For critical actions such as large transfers, authorizations to new addresses, and sensitive contract interactions, AI's capabilities must still be constrained within clearly defined authorization limits through least-privilege principles, human confirmation, pre-execution simulation, and clear explainability.

Especially in genuinely abnormal situations, users must be clearly told "why this is dangerous," "what will happen after execution," and "where exactly the risk lies."

In other words, the significance of AI-powered defense lies in pushing wallets from being passive signing tools to gradually possessing the capability to proactively understand transactions, identify anomalies, and constrain execution.

Final Thoughts

The recent series of wallet security incidents does not mean self-custody has lost its value, nor does it mean users should hand back full control of their assets to centralized platforms.

What they truly remind us of is this: self-custody has never meant inherently secure. It simply returns absolute control of assets to the user.

And protecting that control requires a security system that evolves with the times and continuously upgrades itself. Because security is not a one-time product delivery—it's a long-term, dynamic evolution that requires the collective efforts of users, project teams, and wallet manufacturers.

Attackers can use AI to understand code, users, and environments. Defenders can do the same.

This will be a prolonged "spear versus shield" arms race of escalation.