Crypto Security: Prepare for AI-Powered Attacks14 min read2,732 words

Crypto Security: Prepare for AI-Powered Attacks

AI is lowering the cost of crypto attacks. Learn how teams can secure devices, signing systems and keys against faster, more opportunistic threats.

crypto securityAI-powered crypto attacksDeFi securityNorth Korean crypto hackerscrypto key management
Crypto Security: Prepare for AI-Powered Attacks

Crypto Security Is Entering a Faster, More Dangerous Phase

Crypto teams spend a lot of time on phishing, compromised software, and attempts to steal signing keys. The problem is that these attacks are getting easier to scale. At Solana's Breakpoint conference in London, security leaders warned that AI is lowering the cost - and the know-how needed - to find vulnerabilities, build exploits, and move through an organisation.

That doesn't mean every attacker can now "press a button" and drain a protocol. It does mean less obvious targets and smaller teams may no longer get much protection from the simple fact that attackers are busy or limited. At the same time, North Korean operatives keep using familiar methods - poisoned software, fake job interviews, and attempts to reach transaction-signing systems - to go after crypto firms.

The practical response isn't to buy an AI security product and assume the problem is handled. Teams need to map where an attacker can get a foothold, how they could move from a staff device into production, and what controls stop them from getting to funds. This article breaks down how the threat is changing, the attack patterns described by security leaders, and a way to prioritise defences that actually fits how teams operate.

What does AI change about crypto security?

AI changes the economics of an attack more than it changes the basic logic. Attackers still need an entry point, a way to escalate access, and a path to something valuable. What changes is the speed of testing possibilities, adjusting tactics, and repeating the work across lots of targets.

At Breakpoint, Anchorage Digital co-founder Diogo Mónica said AI systems have recently moved beyond identifying vulnerabilities into building exploits and taking on more of the attack itself. He warned that tools can make complex activity accessible even to people without deep security expertise. Solana Foundation chief information security officer Michael Coates made a similar point: AI has reduced both the cost and the barrier to hacking.

The important nuance is this: faster reconnaissance and better code analysis can help an attacker spot a weak surface sooner, but it doesn't automatically defeat good access controls or turn every bug into a working exploit. Human judgement, infrastructure knowledge, and operational access still matter. The risk is that those steps are increasingly assisted - or automated - so attackers can pursue more targets with less effort.

Why smaller projects should pay attention

A project doesn't need billions in assets to become a target. Attackers can hunt for exposed services, outdated dependencies, vulnerable web apps, employee credentials, or a contractor with access to internal tools. When searching is cheaper, opportunistic targeting becomes more attractive.

That shifts how teams should think about risk. A small group might previously assume it's too obscure for a capable attacker. That assumption gets weaker when automated tools can scan widely and rank "likely" weaknesses. The end result isn't that every project gets attacked - it's that being obscure offers less protection than it used to.

What AI does not replace

Security teams shouldn't treat AI as either magic or irrelevant. It can speed up vulnerability discovery and attack preparation, but it can't cover for every missing prerequisite. A compromised laptop still needs a route into systems that matter. A discovered bug still has to be exploitable in the real environment. A wallet still needs signing authority to move funds.

For defenders, the useful question isn't whether an attack is "AI-powered." It's whether the organisation has reduced the number of reachable weaknesses and limited what an attacker can do after compromising a device or account. Better tooling doesn't remove the value of those controls.

How do attackers move from a device to crypto funds?

Getting from an employee's computer to a protocol's funds is rarely one clean "big exploit." It's usually a chain: compromise a person or device, gain access to internal systems, find a route to privileged services, then abuse signing authority. If you understand that chain, you'll get more leverage than if you focus only on the final wallet or smart contract.

Coates described the Binance breach discussed at the conference as involving familiar steps. A developer downloaded poisoned software after clicking through a sponsored search result. Later, an engineer took a job interview using a work device and downloaded software as part of the process. Those footholds reportedly helped attackers move through the network and reach signing services.

The takeaway is important: attackers don't always need to steal private keys directly. If they can influence a system that's authorised to sign transactions, they may be able to get that system to approve malicious actions. Key material still matters - but so does the boundary around who (or what) can request signatures.

Treat employee devices as part of the production perimeter

A developer workstation can touch source code, cloud accounts, deployment pipelines, chat, email, and administrative dashboards. If one machine connects all of those systems, compromise becomes a bridge between normal corporate activity and critical infrastructure.

That's why device security isn't just an IT issue. Teams should know which devices can reach production, what credentials are stored locally, and how quickly access gets revoked if a device - or an account - is suspected to be compromised. Separating development environments from corporate devices can reduce damage, but it adds operational complexity and can slow workflows.

Protect the signing path, not only the key

Map the chain from "someone makes an action" to "an on-chain transaction happens." Who can propose it? Which service validates it? Who can approve it? Where do signing credentials live? If one account is compromised, can they change a destination address or bypass a review?

A well-designed signing process should make unauthorised transactions hard even after an attacker gets some internal access. In practice, that can mean multiple independent approvals, separating proposal and approval roles, restricting signing services to known workloads, and alerting on unusual requests. The exact implementation varies by protocol and operating model - but the principle is the same: one compromised laptop or login shouldn't be enough to move meaningful funds.

Why are North Korean operations a serious risk?

North Korean hacking groups pose a persistent threat to crypto companies because stolen digital assets can support state priorities. Coates said at Breakpoint that more than $13 billion in crypto had been stolen over the preceding decade, including more than $7.5 billion in the year he was discussing. He attributed about $6 billion to a single Binance breach and said the FBI blamed North Korea for that incident.

These figures came from conference remarks, so they should be read in that context. The broader lesson is more durable than any single number: a well-resourced attacker may combine cyber intrusion with recruitment fraud, social engineering, and access gained through people inside a company.

Recruitment and software supply chains are attack surfaces

Fake interviews aren't just a human-resources problem. Coates described job interviews that persuade engineers to use work devices and download software. A seemingly routine recruitment process can become a way into an organisation's network. Poisoned developer tools and software offered through search results add a related risk: the attacker disguises the malicious download as something legitimate a candidate "needs" for the role.

Companies should be extra cautious when hiring involves unfamiliar software, unusual technical tests, or requests to use a corporate device outside normal procedures. That doesn't mean every unusual interview is malicious. It does mean the process should let teams assess candidates without handing over access to untrusted code.

Insider access can be remote and indirect

Mónica described cases where North Korean operatives use people in the United States to take jobs on their behalf, then use remote keyboards and mice to access a company network. The U.S. Treasury has also sanctioned individuals connected to a North Korean IT worker scheme, according to remarks reported at the conference.

The point isn't to treat every remote worker as suspicious. The point is to verify identity and work arrangements, watch for unusual access patterns, and ensure a person's device can't quietly turn into a proxy for an unknown operator. Anchorage's reported measures included onboarding new staff in person at its U.S. office and using locked-down Chromebooks to connect to remote development machines. Mónica said this approach removed 90% to 95% of its corporate attack surface.

That number reflects one company's setup, not a universal benchmark. A locked-down device model can reduce risk, but it depends on disciplined identity management, endpoint controls, and secure remote environments. Teams should weigh the reduction in exposure against the cost and day-to-day friction of running the model.

Which security controls should crypto teams prioritise?

Start with controls that protect the path to money and privileged access. It's easy to sink time into sophisticated threat detection while leaving basic routes open - weak authentication, unmanaged devices, excessive permissions, or a signing process controlled by too few people.

Coates urged teams to protect keys, including using hardware security keys for logins and properly configured multisignature wallets. He also stressed planning for key rotation and key loss. These measures address different failure modes: phishing-resistant authentication helps protect accounts, multisig reduces reliance on one signer, and a rotation plan limits how long compromised credentials remain useful.

Secure identity and devices first

For employee and administrator accounts, use hardware security keys where possible - especially for email, cloud platforms, code repositories, and identity providers. Email deserves extra attention because password resets and account recovery routes can effectively give someone access to other systems. Keep privileged accounts separate from routine accounts, and review permissions when responsibilities change.

Standardise and manage company devices. Keep systems updated, restrict software installation where feasible, and make it easy to revoke access quickly. A policy only helps if the team can see which devices are connecting - and if they can respond when one is lost, compromised, or no longer under management.

Reduce the chance of a signing mistake

Multisig isn't a complete security programme. Poor signer choices, shared devices, unclear approval rules, or an unsafe recovery process can still undermine it. Define who can propose and approve transactions, how large or unusual transfers get reviewed, and what happens if a signer is unavailable or compromised.

Keep a documented key lifecycle: generation, storage, access, backup, rotation, and destruction. Test recovery procedures before an emergency. If a key is lost, the team should be able to restore operations without improvising under pressure. If a key might be compromised, it should know how to remove its authority safely.

Treat AI agents as privileged users

Mónica warned that connecting email inboxes to AI agents can invite prompt-injection attacks. The core issue is permissions: if an agent can read sensitive messages, follow instructions, and take actions, it can be manipulated by content it encounters. The same risk applies to any automated tool with access to code repositories, cloud consoles, or transaction workflows.

Before connecting an agent to business systems, define what it can read and what it can change. Keep credentials out of its context, require human approval for consequential actions, and log what it does. If the agent doesn't need access to an inbox or signing workflow to do its job, don't grant it.

How can teams build a practical security programme?

A good programme turns security from a pile of tools into a repeatable operating process. Solana Foundation's STRIDE framework, launched in the spring before the conference remarks, was presented by Coates as a free framework with 18 core controls tailored to crypto. He said protocols representing more than 74% of Solana DeFi's total value locked - including 10 of the top 11 - had gone through it.

That adoption number is a useful signal of how widely it's being used, not proof that every participating protocol is secure. A framework helps teams identify and organise controls, but it doesn't remove the work of implementing them properly, testing assumptions, and responding when something fails in practice.

Use a threat model that follows money and access

Start by listing the assets that would cause serious damage if compromised: signing authority, deployment credentials, user data, treasury wallets, cloud accounts, and critical infrastructure. Then map the paths an attacker could take to reach them. Include staff devices, vendors, hiring processes, code dependencies, and automated agents - not just smart contracts.

For each path, write down three things:

  • Entry point: How could an attacker gain initial access?
  • Privilege transition: What would move them closer to production or signing systems?
  • Impact control: What prevents one compromise from becoming a material loss?

This makes prioritisation concrete. A control that blocks a likely route to transaction approval should generally come before a low-probability issue with limited impact.

Test the controls that matter under pressure

Tabletop exercises uncover gaps that documentation can hide. Walk through scenarios like a compromised developer account, a suspicious transaction request, a lost hardware key, or a recruiter sending an unusual software package. Then ask: who detects it, who can revoke access, who communicates with users, and who has authority to pause operations.

Coates also said Solana Foundation was using AI to reverse-engineer newly launched programs and find flaws before attackers. He said it prevented more than $5 million in potential losses during the preceding month. That's a concrete example of using automation defensively - but it should complement code review, testing, and incident response, not replace them.

A focused first-month checklist

If your team doesn't have a clear starting point, use a short sequence instead of rolling out a large security programme with no owner:

  1. List every account and device that can access production, treasury, or signing workflows.
  2. Require phishing-resistant authentication for privileged accounts and remove unnecessary access.
  3. Review how transaction requests are proposed, approved, signed, and monitored.
  4. Check developer tools, software installation practices, and hiring assessments for unsafe downloads.
  5. Define key-loss, key-compromise, and device-compromise procedures, then rehearse them.
  6. Review any AI agent's permissions and remove access it doesn't need.

This won't eliminate risk. It gives the team a solid baseline and a way to find where the next investment reduces the most exposure. The order matters: make access visible first, then restrict it, then test whether those restrictions hold during a real incident.

Conclusion

AI is making parts of cyberattacks cheaper and easier to repeat, while established tactics - poisoned software, fake recruitment, and compromised devices - still work. For crypto firms, the core security question is simple: if an attacker gets one foothold, can they reach the systems that authorise transactions?

Build your defence around that path. Secure identities and devices, limit privileges, separate proposal and approval, protect keys, and rehearse recovery. Frameworks and AI-assisted analysis can help you find gaps, but they only matter once controls are actually implemented and tested in the systems that move funds.

Frequently Asked Questions

No. AI can help discover weaknesses, generate code and accelerate parts of an attack, but a successful breach still depends on the target’s systems, permissions and operational controls. The practical concern is increased speed and scale, not a guarantee that any target can be hacked on demand.
There is no single answer for every company. Common high-impact risks include compromised employee accounts or devices, unsafe software, excessive privileges and weak transaction approval processes. Teams should map how an attacker could reach signing authority and prioritise controls that break that path.
A properly designed multisig can reduce the risk that one compromised key is enough to move funds. It does not prevent every theft: poor signer security, collusion, unsafe approval rules or a compromised signing workflow can still expose assets. Multisig needs a tested key-management and recovery process.
Use AI to assist with tasks such as code analysis or identifying unusual activity, but validate its findings and keep sensitive actions under controlled permissions. Do not give an agent broad access to email, cloud accounts or signing systems unless that access is necessary and safely constrained.

Related articles