In November 2025, Anthropic said a Chinese state-sponsored group pointed Claude Code at roughly thirty organizations and let the model do 80 to 90 percent of the tactical work, with humans stepping in at about four to six decision points per campaign. It got through against only a handful of targets. Still, nobody had to type most of the commands. Strip away the hype and that's the shift in AI cybersecurity right now. Attacks aren't getting stranger. They're getting cheaper, and attackers need fewer skills. In crypto, where a signed transaction can't be pulled back, that matters more than almost anywhere else.
What counts as an AI-related cyber incident, and what doesn't
The label gets stretched a lot. An incident qualifies when attackers used AI inside the operation itself, say by cloning a voice or running part of the intrusion. It doesn't qualify just because the victim is an AI company. Bybit lost about $1.5 billion in February 2025 and the FBI blamed North Korea, but the public notice describes no AI role, so it's an ordinary heist, just a huge one. Plenty of the cases below sit in the gray middle, where AI clearly helped and nobody can prove how much.
Phishing, voice clones and fake executives
Social engineering is where AI pays off first, since text and faces are exactly what these models do well. Bad grammar used to give phishing away. Nobody can lean on that test anymore, and ai phishing attacks now arrive polished and personal. Business email compromise follows the same pattern. A fake invoice thread used to be clumsy. Now it can mimic a supplier's tone and reference real projects, and the usual gut check no longer holds.
The video call that cost $25 million
The best-documented case predates 2025 but still works as the template. In early 2024, a finance employee at Arup's Hong Kong office joined a video call where the CFO and several colleagues were all deepfakes, then sent roughly $25 million across 15 transfers. Arup said none of its systems were breached. The fraud ran entirely on trust. Deepfake scams like this one worry finance teams more than any exploit does.
Cloned voices of senior officials
In May 2025 the FBI warned that attackers were sending texts and voice messages, some using AI-generated voices, to pose as senior US officials and reach their contacts. The advice is almost comically low tech: hang up, find a real number yourself, call back. Families can agree on a code word, and so can finance teams.
Malware and automated hacking
The overlap between ai and cybersecurity gets oversold here, and the primary reports admit as much.
Malware that asks a chatbot for help
Google's threat intelligence group described PROMPTFLUX, a script that asks Gemini for fresh evasion code about once an hour. Google called it experimental and noted it can't compromise a victim network yet. PROMPTSTEAL was further along. Russia's APT28 used it in Ukraine, and it queries a model for commands that collect documents, which Google called its first observation of ai malware querying an LLM in live operations. The same report says the underground market for AI tools matured in 2025.
Agents running the intrusion
The Anthropic campaign above is the clearest example of ai hacking at scale, and it had limits: Claude sometimes invented credentials and passed off public data as secrets. A second case, from Anthropic's August 2025 report, involved one actor using Claude Code against at least 17 organizations, including healthcare and emergency services. The model sorted through stolen financial data to help set ransom demands that sometimes topped $500,000.
Is AI really lowering the bar?
Yes, mostly. That same August report describes a criminal with only basic coding skills who used Claude to build ransomware variants and sell them for $400 to $1,200 on forums. Skill hasn't stopped mattering, because models still invent credentials and miss obvious things. But the gap between wanting to attack and being able to has narrowed. Small companies that never expected to be targeted now face real AI security threats.
Where this hits crypto hardest
Crypto gets hit twice. It holds money that moves in one irreversible step, and many teams are small enough to lack a security desk.
Fake meetings and North Korean crews
Mandiant described a February 2026 intrusion at a crypto-sector fintech firm, tied to the North Korea-nexus group UNC1069. A hijacked Telegram account led the victim to a spoofed Zoom call, where a fake audio problem pushed them to run "troubleshooting" commands. Seven malware families followed. The victim thought one participant looked like a deepfake, but Mandiant couldn't confirm AI use here. The confirmed part is the social engineering, which is worth keeping apart from the deepfake claim. Google's earlier report did document the same group using Gemini for crypto-themed lures.
Scams at industrial scale
Chainalysis estimates $17 billion was stolen in crypto scams in 2025, and its on-chain data shows impersonation scams up more than 1,400% on 2024. Operations with on-chain links to AI vendors averaged $3.2 million each, against $719,000 for the rest. Better-funded crews may simply buy more tools, so read that as correlation. The FBI's 2025 report, released in April, put crypto-related losses at $11.37 billion and flagged roughly 22,000 AI-linked complaints worth about $893 million, so ai fraud now has its own line in the federal ledger.
Fake support accounts and invented token promotions both need believable text and faces, which is exactly what ai scams produce cheaply. Address poisoning is a cousin. In December 2025, one trader sent almost $50 million in USDT to a look-alike address copied from their own history. Reporting doesn't tie AI to it. It's the sort of cheap, repeatable trick that automation rewards, though.
Smart contract bugs in a simulator
Anthropic's Frontier Red Team tested agents against 405 contracts that were really exploited between 2020 and 2025. Models produced working exploits for 207 of them in simulation, worth about $550 million. On 2,849 fresh contracts, two models found two new flaws, but the run cost $3,476 in API fees against $3,694 in exploit value. Nothing touched a live chain. Simulated exploit revenue on the 2025 problems roughly doubled every 1.3 months. That's one benchmark's trend line, and it only covers test funds.
Defenders have the same tools
The arms race runs both ways. In July 2025, Google said its Big Sleep agent found a SQLite flaw that only attackers knew about, which let Google stop exploitation before it landed. It believes that was a first. Exchanges and banks also use models for anomaly monitoring and fraud screening. Auditors can point agents at a contract before launch. The weak point is training data: a detector tuned on yesterday's fakes can miss tomorrow's. The race feels lopsided in places. Attackers only need one message to land, while defenders have to catch every one, and a fraud model that blocks too many honest customers gets switched off by the business.
Warning signs and practical steps
Almost every case above leans on urgency and a move away from official channels.
For individual users
Be wary of anyone who contacts you first and pushes you to hurry or switch apps. Call back on a number you already had. Never run commands a meeting host pastes into chat. Compare a full wallet address, not just the first and last characters, and confirm it on a hardware wallet screen. Agree on a code word with family.
For exchanges, projects and staff
No transfer should rest on a single call or video meeting. Require a second approver on a separate channel and use phishing-resistant hardware keys. Run fake-executive drills for staff. Treat selfie and video KYC as beatable and layer other checks on top. Watch for support accounts that copy your branding, since users will report them to you first, and keep an incident plan that names who can freeze withdrawals at 3 a.m. If you deploy AI agents, give them narrow permissions, because an agent with broad access is a target too.
Do the usual basics still hold up?
Largely, yes. Multi-factor authentication and patching still block much of this. Arup and the impersonated officials were failures of process, where one person could act on a convincing request. What breaks are controls that rely on something sounding or looking right. Companies preparing for this should rewrite approval rules so identity gets checked through a second channel, not judged by how convincing someone seems, and test those rules the way they test backups.
The ai cybersecurity threats worth the most attention next are agent-run intrusions and deepfake meetings aimed at crypto staff. Small projects without a security team come right after. Speed is the common thread. If exploit-building keeps getting quicker, the gap between a bug becoming known and being abused shrinks, and most ai cyber threats will be decided by who patches first.
Matvii Diadkov
Disclaimer: All materials on this site are for informational purposes only. None of the material should be interpreted as investment advice. Cryptocurrency investments involve significant risk of loss. Please note that, despite the nature of much of the material created and hosted on this website, HODL FM operates as a media and informational platform, not a provider of financial advisory services. The opinions of authors and other contributors are their own and should not be taken as financial advice. If you require advice, HODL FM strongly recommends contacting a qualified industry professional.
