State-Sponsored Hackers Exploit AI in Cyberattacks: Google

State-sponsored hackers are already using AI to speed up cyberattacks, and Google says groups linked to Iran, North Korea, China, and Russia have leaned on large language models like Gemini to research targets, write convincing phishing lures, and even support malware development. If you’re in crypto or you run anything on-chain, you can’t treat this as “future risk.” Instead, you should assume attackers can iterate faster than your team, and you should harden your wallets, your comms, and your cloud identities now—because AI makes old scams scale, and it makes new ones harder to spot.

Google’s Threat Intelligence Group (GTIG) has been tracking how government-backed actors fold AI into the attack lifecycle—reconnaissance, social engineering, and tooling. In other words, they aren’t using AI as a magic “hack button.” Rather, they’re using it like a force multiplier: faster research, better language, more believable pretexts, and quicker experimentation. Meanwhile, crypto firms, DeFi protocols, and Web3 communities remain high-value targets because money moves quickly, reversals are rare, and operational security varies wildly.

In this post, I’ll walk you through what Google reported, how these tactics map to crypto-specific risks, and what you can do today to reduce the odds that you—or your team—become the next headline. I’ll also keep it practical: you’ll see where AI actually helps attackers, where it doesn’t, and how to build defenses that still work when adversaries can generate a thousand variants of a lure before lunch.

What Google’s report really says: AI is a multiplier, not a replacement

Google’s GTIG describes a clear trend: state-backed groups are integrating AI into their workflows to accelerate tasks that used to consume time and specialist effort. That includes drafting phishing messages, translating content for specific regions, summarizing technical material, and brainstorming ways to approach a target. Because of this, the “cost per attempt” drops, and attackers can test more angles until something sticks.

Importantly, GTIG’s framing aligns with how most mature security teams see AI in the wild: it doesn’t replace core tradecraft, but it does reduce friction. For example, an operator who struggles with English can now produce polished messages that sound like a native speaker. Likewise, a team can quickly summarize a target’s tech stack from public breadcrumbs and then craft a pretext that fits. As a result, your usual “spot the bad grammar” instincts won’t save you.

Google’s own public AI work provides context on the models being referenced, including Gemini. You can read more about Gemini directly from Google DeepMind here: https://deepmind.google/technologies/gemini/. Separately, GTIG regularly publishes threat intelligence updates through Google Cloud’s threat intel channel, which helps you track the bigger pattern rather than chasing one-off incidents: https://cloud.google.com/blog/topics/threat-intelligence.

So what’s the key takeaway for crypto and blockchain teams? Attackers can now scale “human-sounding” interaction. They can also tailor lures to your niche—staking, governance proposals, audits, bug bounties, exchange listings—without needing a domain expert on staff. Therefore, you and I’ve to shift from “can I spot a scam?” to “can my systems withstand a scam that looks real?”

Where AI shows up in the attack lifecycle

GTIG points to AI use across multiple stages. First, reconnaissance: attackers can quickly gather and synthesize open-source intelligence (OSINT) on employees, vendors, and tooling. Next, social engineering: they can generate targeted messages, role-play conversations, and refine tone for specific personas. Finally, in some cases, they can assist with scripting, debugging, or iterating on malware components—although experienced operators still need real expertise to build reliable payloads.

However, the biggest change isn’t that AI creates brand-new classes of exploits. Instead, it makes existing techniques cheaper and faster. That’s bad news for crypto because the ecosystem already runs on speed: listings, airdrops, governance votes, and incident response all happen under time pressure. When attackers can match your pace—or exceed it—you can’t rely on “we’ll notice something off.” You need controls that don’t depend on vibes.

Why crypto and blockchain teams are especially exposed

Crypto is a perfect storm for adversaries. You’ve got irreversible transactions, public social graphs, and communities that coordinate in public channels. Meanwhile, many teams operate globally, hire quickly, and rely on contractors. That’s not “bad,” but it does expand the attack surface. On top of that, crypto brands often place a premium on being responsive—so if someone pings you about a “critical vulnerability,” you’ll feel pressure to act fast.

State-sponsored actors historically target defense, government, and critical infrastructure. Yet they also target revenue streams, sanctions evasion, and intelligence collection. That’s why crypto firms, exchanges, OTC desks, bridge operators, and even prominent individual traders can end up on the menu. If you’re thinking, “I’m too small,” I wouldn’t bet on it. Attackers don’t need to compromise the biggest protocol if they can compromise a mid-level contributor with access to a multisig, a CI pipeline, or a community announcement channel.

North Korean-linked activity, in particular, has been repeatedly associated with crypto theft campaigns over the years. If you want a high-level view of how the U.S. government frames DPRK cyber activity, CISA maintains ongoing advisories and guidance you can track: https://www.cisa.gov/. Of course, you don’t need to be a nation-state target to get hit; you just need to be adjacent to money, access, or influence.

What’s more, AI makes it easier to craft “crypto-native” pretexts. Attackers can convincingly impersonate:

• Audit firms requesting “one last build” or “verification of the final report.”
• Wallet providers asking you to “confirm integration steps.”
• DAO contributors pushing an “urgent governance hotfix.”
• Exchange listing teams asking for “updated token metadata.”
• Influencers offering “partnership details” with a document link.

Because these scenarios feel normal, you and your team might not slow down. So, the defense has to be procedural and technical, not just educational.

AI-powered phishing hits crypto harder than most industries

Phishing is old, but AI upgrades it. Attackers can A/B test subject lines, generate localized slang, and mimic internal writing styles. They can also create multi-step narratives that feel like real operations work. For example, you might get a “calendar invite” followed by a “quick doc” followed by a “can you run this script?” Each step seems minor. However, the chain leads to credential theft, session hijacking, or malware execution.

In crypto, one stolen identity can unlock a lot: admin access to Discord or Telegram, rights to push website updates, access to a GitHub org, or the ability to propose a multisig transaction. Therefore, phishing doesn’t need to steal private keys directly. It just needs to steal the right account at the right time.

What “AI-assisted malware development” means in practice

When people hear “AI helps build malware,” they often imagine an LLM generating a fully functional, stealthy implant on demand. That’s not how it usually works. Instead, AI tends to help with small but meaningful tasks: writing boilerplate code, explaining APIs, converting code between languages, fixing bugs, and generating scripts for automation. In other words, it reduces the time between idea and execution.

For defenders, that means you’ll see more variants and faster iteration. A payload that used to take days to retool might now take hours. And, attackers can generate more “supporting artifacts,” like fake documentation, fake invoices, or fake internal tickets, which makes the overall campaign more believable.

Still, AI has limits. It can hallucinate, it can produce insecure code, and it can make mistakes that a skilled analyst will catch. However, you can’t count on attacker incompetence. If a state-backed team combines skilled operators with AI acceleration, you should assume they’ll produce high-quality lures and reasonably competent tooling.

To ground your understanding in broader best practices, it helps to align with widely accepted frameworks. MITRE ATT&CK remains one of the most useful public knowledge bases for mapping adversary behavior to defenses: https://attack.mitre.org/. If you map your controls to ATT&CK techniques, you’ll build resilience that holds up even when the attacker writes cleaner emails.

Crypto-specific malware goals you should expect

In the crypto niche, malware often aims for a few high-impact outcomes. First, credential theft: browser-stored passwords, session tokens, and SSO cookies. Second, developer compromise: SSH keys, GitHub tokens, CI secrets, and signing keys. Third, transaction manipulation: clipboard hijacking, address substitution, and tampering with build artifacts. Finally, surveillance: monitoring comms to time a social-engineering strike, especially around treasury movements.

Because of that, you should treat endpoints and developer laptops as critical infrastructure. If you’re thinking “we’re decentralized,” remember: your contributors still use devices, and your protocol still depends on software supply chains. Decentralization doesn’t eliminate trust; it redistributes it.

Defensive playbook: what you can do this week (even if you’re small)

You don’t need a giant SOC to reduce risk. You do need discipline. If you’re running a protocol, an exchange, an NFT platform, or even just a community with a treasury, you can implement controls that make AI-enhanced attacks less effective. Below are steps I’d prioritize, in order, because they reduce the most common failure modes.

1) Lock down identities (because phishing targets accounts first)

Start with identity, because most modern compromises begin with stolen credentials or hijacked sessions. So, do these now:

• Use phishing-resistant MFA (FIDO2/security keys) for email, GitHub, cloud consoles, and admin panels. SMS won’t cut it.
• Enforce SSO where possible, and require device posture checks for privileged access.
• Turn on conditional access rules (geo, device, risk signals) and alert on impossible travel.
• Separate admin accounts from daily accounts. You shouldn’t browse the web on an admin session.

What’s more, make sure you can revoke sessions quickly. If an attacker steals a cookie, password resets alone won’t help. Therefore, learn how to invalidate active sessions across your identity provider and critical SaaS tools.

2) Harden your comms channels (Discord and Telegram aren’t “just community”)

Crypto teams live in chat. That’s fine, but you need guardrails:

• Restrict who can post announcements, pin messages, or change server settings.
• Require MFA for moderators and admins, and audit roles monthly.
• Use separate accounts for moderation versus personal use.
• Publish an “official links” page and never share sensitive links only in chat.

On top of that, train your community to distrust urgency. Attackers love “last chance,” “critical exploit,” and “airdrop ending.” If you normalize verification steps, you’ll reduce panic clicks.

3) Protect the software supply chain (because that’s how big losses happen)

If you ship code, you’re a supply-chain target. So:

• Require code review and branch protection for production repos.
• Use signed commits and signed releases where feasible.
• Rotate secrets, store them in a proper secrets manager, and eliminate long-lived tokens.
• Lock down CI/CD permissions; don’t let every workflow access every secret.

Also, watch for “helpful” PRs from new contributors. They might look legitimate, and AI can make them look even more polished. As a result, you should treat new contributor code as untrusted until proven otherwise.

How AI changes incident response for crypto teams

Even with strong defenses, you should plan for compromise. AI changes incident response because attacks can move faster and branch into multiple parallel attempts. While you investigate one lure, the attacker might launch three more variants at your other admins. Therefore, your response needs to be decisive and rehearsed.

Here’s what I recommend you document and practice:

• A “break glass” process to freeze treasury movements and pause high-risk operations.
• A contact tree that includes exchange partners, bridge partners, domain registrars, and hosting providers.
• A rapid credential reset plan: email, SSO, GitHub, cloud, Discord/Telegram, and DNS.
• Log retention and access: you can’t investigate what you didn’t store.

Also, you should assume attackers will exploit confusion. So, establish an out-of-band verification method (for example, a pre-shared phrase or a verified call-back number) for any request involving keys, deployments, or announcements. If you and I only do one thing differently after reading Google’s report, it should be that: verify sensitive requests outside the channel where the request arrived.

On-chain realities: containment looks different when transactions are final

In traditional finance, you might reverse a transfer. On-chain, you usually can’t. So, containment often means preventing the next transaction, not undoing the last one. That’s why multisig policies, timelocks, and spending limits matter. If an attacker compromises one signer, they shouldn’t be able to drain everything immediately.

If you operate a treasury, consider layered controls:

• Multisig with distributed signers and strict opsec requirements.
• Transaction simulation and human-readable signing policies.
• Timelocks for large transfers and contract upgrades.
• Separate hot funds (ops) from cold funds (reserves).

These controls don’t stop every attack. However, they buy you time, and time is the one resource attackers can’t easily steal—unless you give it away by rushing.

What you should watch next: AI, deepfakes, and “trust layer” attacks

Today, GTIG emphasizes AI for research and phishing. Next, you should expect more “trust layer” attacks: deepfake voice calls to signers, fake Zoom meetings with synthetic video, and cloned writing styles that mimic your CEO or lead dev. Because crypto teams already work remotely, attackers can exploit that norm. If you’ve never met a teammate in person, a convincing video call can feel “good enough.”

So, we need stronger verification rituals. For example, you can require that any urgent request involving deployments or funds must be confirmed via two independent channels. Likewise, you can require that signers confirm transaction intent in a separate, authenticated system, not just in chat. These habits feel slow at first. However, they quickly become muscle memory, and they’ll save you when the lure looks perfect.

Finally, keep your threat intel intake realistic. You don’t need to read everything, but you do need a few reliable sources and a cadence. Track vendor reports, follow major advisories, and review your own telemetry. If you do that, you’ll spot patterns earlier—and you won’t be surprised when attackers reuse a playbook that worked elsewhere.

FAQ

Are state-sponsored hackers really targeting crypto, or is this mostly about governments?

They’re targeting both. Even when a campaign’s primary goal is intelligence collection, crypto companies can provide access to money, infrastructure, or strategic insight. Also, some state-linked groups have a documented history of crypto theft, so you shouldn’t assume you’re “out of scope” just because you’re not a defense contractor.

Does AI mean phishing emails are impossible to detect now?

No, but they’re harder to detect by “feel.” AI can remove obvious red flags like poor grammar or awkward phrasing. Therefore, you should rely more on authentication (DMARC/SPF/DKIM), access controls, and verification processes than on human intuition alone.

What’s the single best control to reduce AI-driven social engineering risk?

Phishing-resistant MFA (security keys/FIDO2) on your most critical accounts is the best immediate win. It won’t solve everything, but it dramatically reduces the impact of credential theft. Pair it with session revocation and least-privilege access, and you’ll cut off many common attack paths.

How can a small DeFi team improve security without hiring a full SOC?

Focus on high-tap into basics: lock down identity, protect admin roles in chat platforms, enforce code review and branch protection, and segment treasury funds with multisig and timelocks. Also, write a one-page incident plan so you’re not improvising under pressure.

Where can I learn more about the tactics attackers use?

Start with MITRE ATT&CK for a structured view of techniques and mitigations: https://attack.mitre.org/. For ongoing public advisories and practical guidance, you can also monitor CISA: https://www.cisa.gov/. These resources won’t replace internal logging and reviews, but they’ll help you prioritize defenses that work against real-world tradecraft.

You Might Also Like: Intuit, Uber, and State Farm trial enterprise AI agents, Google DeepMind Unveils AlphaGenome: A Unified Sequence-to-Function Model Using Hybrid Transformers and U-Nets to Decode the Human Genome