In a mid-October interview with CNN, Microsoft co-founder Bill Gates predicted that the most dramatic changes in the coming year would stem from cyberattacks.
He noted that developers can disable the safety mechanisms built into AI models, and the real danger is their exploitation by individuals with malicious intent.
Outlets covering Gates’s comments cite data from CrowdStrike showing an 89% year-over-year increase in AI-powered cyberattacks in 2025.
The danger begins with a conversation
Vugar Usi, head of major crypto exchange MEXC, suggests viewing this issue from the perspective of an average user who interacts with an AI assistant daily.
He argues that the primary risk isn’t obvious targets like passwords or seed phrases, which most cryptocurrency users have already learned to protect, but information that may seem harmless in isolation.
Usi cites an investor who, over several conversations, might mention their dog’s name, discuss losses on perpetual futures, ask about withdrawing funds from a specific exchange, and talk about their mortgage or their child’s education expenses.
While no single detail is alarming, taken together, they could reveal to a fraudster exactly what kind of narrative would compel that individual to act urgently.
Modern fraud no longer needs to begin with a generic “Dear Customer” greeting, Usi points out.
Instead, a message might open with details that make it seem the sender knows an actual account and a specific, genuine issue.
Data leaks make this easier: names, email addresses, and phone numbers have long since ended up in compromised databases.
Furthermore, the history of one’s interactions with an AI can provide missing context, stitching together disparate pieces of information into a detailed financial profile.
Consequently, Usi notes, risk is increasingly determined not by the sensitivity of any single piece of data, but by what can be inferred from the combination of such details.
The scale of the habit
This practice has become widespread. According to a 2024 Experian survey cited by Usi, nearly half of American consumers have used (or considered using) generative AI to manage their personal finances. This figure rises to 67% among Gen Z and 62% among millennials.
Within the crypto community, the willingness to trust AI goes even further: a CoinGecko survey found that 35.9% of respondents would be comfortable entrusting 60% or more of their portfolio to an AI agent, with roughly one in seven willing to hand over their entire portfolio.
A 2024 study by Cisco found that 30% of generative AI users have entered personal or confidential information into these systems: 29% shared financial details, and 27% provided account numbers.
An analysis by Harmonic Security, covering over 22 million prompts and file uploads in workplace environments in 2025, identified more than 579,000 instances of sensitive data exposure.
In a separate study, the company found sensitive material in approximately 4.4% of prompts and 22% of uploaded files.
According to Usi, files are a particularly overlooked risk.
For instance, a tax return stripped of a name and Social Security number does not become anonymous; details regarding one’s employer, income, and investments still point directly to the owner.
Furthermore, AI tools generally do not require account numbers or precise balance figures to perform portfolio analysis.
Usi also warns against assuming that the services themselves are inherently secure.
In January 2025, researchers at Wiz discovered an exposed DeepSeek database containing over a million rows of logs, including chat histories and secret keys, that required no authentication to access; the vulnerability was closed after the company was notified.
Usi notes that AI assistants are not necessarily more dangerous than other services — banks and even government agencies have all experienced data breaches — but the principle of data minimization is paramount.
Information that a user never shares cannot be stolen from a third-party database, which becomes increasingly critical as AI assistants evolve from tools for one-off queries into persistent companions that remember user preferences and help inform personal decision-making.
For over a decade, the crypto industry has urged users to safeguard their private keys and seed phrases — a rule that remains as vital as ever.
However, in the age of AI, the head of MEXC suggests that the classic maxim, “Not your keys, not your coins,” should be supplemented by a new one: if a machine doesn’t need to know something, don’t tell it.
The post MEXC CEO warns users against over-sharing with AI appeared first on Invezz
