Artificial intelligence is becoming a bigger part of cryptocurrency security, helping exchanges analyze behavioral signals, investigate suspicious transactions, and identify threats at a scale difficult for human teams to match. But the same tools are also reducing the cost of sophisticated attacks.
The recent Coldcard hardware-wallet incident illustrates that tension. A years-old firmware integration flaw weakened the randomness used to generate some wallet seeds, while manufacturer Coinkite said its own AI-assisted code review had failed to identify the issue. The company has also said it must assume attackers could have used AI to inspect historical open-source firmware.
In this TNGlobal Q&A, Jeff Ko, Chief Analyst at CoinEx, talks about what the incident means for self-custody, how exchanges are using AI defensively, the risks of increasingly autonomous financial agents, and why human judgment remains important even as security becomes more automated.
How has your background shaped the way you think about cybersecurity, AI and crypto?
My background gives me three different perspectives on the same problem. I started in private banking, where protecting client assets, managing downside risk and maintaining institutional trust were central. I then founded a data-visualization startup, which made me think more deeply about how technology and data influence human decision-making. In crypto research and investment, finance, software and cybersecurity are essentially inseparable.
That makes me see cybersecurity less as a pure IT issue and more as a combination of financial risk and human behavior.
Chainalysis estimated that more than $3.4 billion in cryptocurrency was stolen in 2025, with the Bybit incident alone accounting for about $1.5 billion. The top three service hacks represented 69 percent of service-related losses. That illustrates the importance of tail risk: one failure in custody, signing or key management can overwhelm years of incremental security improvements.
AI adds another dimension. It can analyze information at a scale humans cannot, but it can also make complicated financial actions feel deceptively simple. Security increasingly depends not only on whether the technology works, but whether users understand what it is doing on their behalf.
Where is AI actually working in crypto risk detection today, and where does it still fall short?
The most successful deployments are narrower than some of the marketing suggests.
One area is behavioral detection for account takeover, using signals such as device fingerprints, typing patterns, session behavior and navigation. Another is market surveillance, where models can identify patterns such as wash trading, spoofing and coordinated manipulation across large transaction networks.
A third area is investigation support. AI can summarize wallet histories, consolidate related alerts and help analysts prepare case narratives. That may not sound as dramatic as autonomous threat detection, but reducing investigation backlogs can have meaningful operational value.
The limitations are equally important. Illicit activity has a low base rate, so even accurate models can generate large numbers of false positives. Labels also arrive late: an exchange may only learn weeks later that a withdrawal was connected to fraud.
The right architecture, in my view, is AI for detection and prioritization, deterministic rules for clearly defined high-risk actions, and humans for ambiguous or high-impact decisions. The goal should not be maximizing the percentage of decisions made by AI. It should be allocating human attention more effectively.
What does the Coldcard incident tell us about AI and cybersecurity risk?
The Coldcard incident is particularly interesting because AI sits on both sides of the security equation.
The underlying problem was not that a hardware random-number generator failed and software automatically replaced it. A build and link integration error caused wallet seed generation to resolve to a general-purpose software pseudorandom number generator rather than the intended hardware source.
Coinkite estimated an effective search space of roughly 40 bits for affected older models and around 72 bits for later devices, compared with its intended 128-bit security target.
The more important AI lesson is that Coinkite said it had used a leading AI model to review its code only weeks earlier and that the model had not identified the problem. Because the firmware was open source, the company said it had to assume an attacker could also have used AI to inspect historical code.
That does not prove AI found this particular vulnerability. But it illustrates a broader change: old security debt is becoming easier to rediscover. Attackers can increasingly re-audit years of open-source software at low marginal cost.
AI is democratizing code review — for defenders and attackers alike.
Does this change the case for self-custody?
I don’t think the lesson is that self-custody has failed. The more important lesson is that self-custody relocates risk rather than eliminating it.
Moving assets away from an exchange can reduce exposure to exchange insolvency or account compromise. But the user then depends on hardware, firmware, random-number generation, backups, supply chains, and their own operational security.
The Coldcard case is important because users could have followed normal security procedures while still relying on a seed generated with weaker-than-intended entropy.
The discussion therefore needs to go beyond “not your keys, not your coins.” We should also ask how those keys were created, where they are stored, and how many independent failures would be required before someone could move the assets.
For sophisticated holders, that strengthens the case for approaches such as multisignature arrangements and diversified signing infrastructure. Owning a hardware wallet should not be confused with eliminating technology or counterparty risk.
Is AI making crypto easier for mainstream users?
Yes. Crypto has historically had as much of a user-experience problem as a technology problem.
Users have needed to understand addresses, gas fees, networks, bridges, smart-contract approvals and seed phrases. AI can potentially abstract much of that complexity into natural-language instructions.
There is also a second form of adoption emerging: machines themselves becoming crypto users.
Coinbase reported that its x402 machine-payment protocol had processed more than 100 million payments by the first quarter of 2026, with more than 99 percent of x402 transactions using USDC. It also reported that more than 90 percent of measured onchain agentic stablecoin transaction volume was occurring on Base.
That suggests crypto’s next wave of users may include software agents acting for humans.
The challenge is that simplifying the interface does not simplify the underlying risk. An irreversible blockchain transaction can start to feel like sending a text message.
That is why I favor progressive autonomy: allow AI to research and recommend freely, but expand financial authority gradually through spending limits, contract simulation, whitelists and additional approval for unusual transactions.
How might exchanges evolve as agentic AI becomes more common?
Exchanges could gradually shift from destinations that humans log into toward financial infrastructure that machines access continuously.
We are already seeing exchanges and wallet infrastructure expose more machine-readable interfaces, command-line tools and Model Context Protocol integrations. Agent-focused wallets are also introducing controls such as per-transaction limits, session spending caps, transaction screening and isolation of private keys from the AI model itself.
That changes how exchanges may compete.
Today, platforms are largely compared on liquidity, fees, product range and user experience. In an agentic environment, additional factors become important: API reliability, machine-readable liquidity, execution quality, permission architecture and agent security.
An agent could eventually choose dynamically between a centralized exchange, decentralized exchange or another liquidity venue based on price, market depth, latency and risk.
Where should humans remain in the loop?
AI should primarily be treated as a decision-support layer when decisions directly affect customer assets.
Its strengths are speed and scale: screening transactions, identifying unusual behavioral patterns and prioritizing cases for investigation.
Humans should remain responsible where context and judgment matter most — complex investigations, exceptions, model calibration, account restrictions, withdrawal blocks and other high-impact or irreversible actions.
I think of the ideal model as layered: AI flags and recommends, deterministic controls establish boundaries, and humans make the final judgment when the potential consequences are material.
AI provides speed, consistency, and scale. Humans provide context, judgment and accountability.
How is attackers’ use of AI changing the threat model?
The biggest structural change is that AI removes much of the traditional trade-off between scale and personalization.
Historically, mass phishing campaigns tended to be generic, while highly targeted social engineering required considerable manual research. AI allows attackers to scrape a victim’s public footprint, generate convincing messages in different languages and styles, clone voices and maintain many tailored conversations simultaneously.
The same change is happening technically. AI agents can assist with reconnaissance, code review, vulnerability discovery and the generation of possible exploit paths.
The weakest point may therefore no longer be a conventional software vulnerability. It may be a convincing interaction that persuades a legitimate user or employee to perform a harmful action — signing a transaction, approving a withdrawal, resetting credentials or revealing sensitive information.
Defense therefore has to extend beyond phishing filters toward behavioral monitoring, transaction-level controls, stronger authentication and continuous verification of user intent.
What should crypto security teams watch over the next 12 months?
One of the biggest shifts will be AI turning old code into a new attack surface.
I expect more incidents where a vulnerability was introduced years earlier rather than in the latest release. As AI models become better at reviewing large historical codebases, attackers can continuously revisit randomness generation, key-derivation logic, bridge verification code, and abandoned dependencies that previously escaped scrutiny.
Beyond that, three areas will matter.
First, exchanges, blockchain networks, stablecoin issuers and analytics providers will need faster sharing of risk signals so stolen funds can be identified and, where possible, intercepted before being dispersed.
Second, behavioral biometrics will become more important for account security, supplementing passwords and devices with session and transaction behavior.
Third, the AI-versus-AI contest around identity will intensify. As synthetic identities, voice cloning and generated video improve, static identity checks will become less reliable, pushing platforms toward dynamic liveness tests, multimodal verification and more interactive authentication.
The broader shift is toward continuous security. Exchanges and wallets will increasingly need to reassess whether a user, device, transaction or piece of code is safe as new intelligence becomes available.
Editor’s note: This Q&A has been lightly edited for clarity, length and factual precision. TNGlobal checked key publicly available factual claims and, where necessary, omitted or qualified a small number of figures that could not be independently corroborated. The substance of the interviewee’s responses has been preserved.
Share your perspective: TNGlobal welcomes contributed insights and expert commentary from across Asia’s technology and innovation ecosystem. Submit a contribution for editorial consideration, or explore more conversations in our TNGlobal INSIDER and TNGlobal Q&A and Interviews archive.
What 1,000 EV chargers reveal about making Singapore’s older condos EV-ready [Q&A]