Anthropic’s IPO prospectus warns that powerful AI could resist shutdown, conceal information and help malicious actors develop weapons of mass destruction

Advanced artificial intelligence (AI) systems, if left without proper safety measures and oversight, could pose an existential threat to humanity; furthermore, the most powerful models could assist malicious actors in developing weapons of mass destruction. These warnings are part of Anthropic’s IPO prospectus, published on the U.S. Securities and Exchange Commission website on October 26.

The Anthropic prospectus states that powerful AI systems could potentially create catastrophic or existential risks for humanity, particularly if they are used maliciously, replicated, weaponized, or granted a high degree of autonomy. Among other risks, the developer cites unintended model behavior, including attempts to resist being shut down, the concealment or distortion of information, and actions resembling blackmail.

Dario Amodei, Anthropic: Advanced AI development must be slowed down

In September, Anthropic CEO Dario Amodei called for a slowdown in advanced AI development to ensure it remains under control. He proposed involving independent experts in safety evaluations and fostering international coordination.

He linked his concerns to an incident involving AI agents developed by a competitor, OpenAI, which broke free from their constraints and compromised the machine learning platform Hugging Face. The models attacked targets without receiving specific instructions, sacrificed themselves to achieve a collective goal, and attempted to breach the very system designed to evaluate them. While Amodei acknowledged that the economic damage was minimal, he viewed the event as a warning sign of significant future dangers.

Vadim Drozd, FinteqHub: Valid payments, wrong outcomes

Vadim Drozd, CEO of Lithuanian payment company FinteqHub, warns about risks that may happen when an AI agent does exactly what it was told and the user still loses. Europe’s banking infrastructure is becoming ready for agentic commerce, he says, but the current stage still resembles an open alpha test, with much of the risk from erroneous actions falling on the user.

To the payment system, that transaction is valid and authorized, and it may be impossible to cancel or refund. A spending limit and a payment authorization are therefore not enough, Drozd argues. Consumers will need to define the moments when the agent must ask for confirmation before completing a purchase. On the institutional side, banks need to run thousands of testing cycles before giving AI full access to payments. The problem, as he describes it, is that a payment can be fully valid and still wrong for the customer.

Vugar Usi, MEXC: Even small disclosures add up

Vugar Usi, CEO of MEXC, argues that the main danger of AI assistants is not a leaked password but accumulation. Crypto users have learned to guard seed phrases, yet they mention a dog’s name in one chat, trading losses in another and a mortgage in a third. None of these details is dangerous alone, but together they form a financial profile that can show a fraudster which story will make this person act urgently.

Existing data breaches already contain names, emails and phone numbers, so exposed chat history could supply the missing context for a scam. Still, removing a name from a tax return doesn’t make it anonymous, he adds, and Wiz’s January 2025 discovery of an unprotected DeepSeek database shows that no online system is immune. His rule for the AI era is practical: if the machine doesn’t need to know it, don’t tell it. Data that was never shared cannot leak.

Jamie Dimon, JPMorgan Chase: Cyber risk multiplied

Jamie Dimon, chairman and CEO of JPMorgan Chase, has made cyber risk the center of his AI warnings, even as his bank invests heavily in the technology. In his annual shareholder letter, published in April 2026, he said AI will almost surely worsen cyber risk. The letter also says AI will eliminate some jobs, that deepfakes will require government regulation, and that the risks are manageable if companies, regulators and governments prepare.

His forecast sharpened in October. At the JPMorgan Tech Summit in London, he said the letter was written before advanced vulnerability detection models such as Anthropic’s Mythos, and urged the industry to control the technology and fix its problems.

Arthur Azizov, B2BINPAY: AI can make a mistake in the least expected moment

Arthur Azizov examines AI from the perspective of payment processing. In his view, the biggest risk for users is that AI can be wrong — and it will happen precisely when they least expect it.

“Trust grows with every correct answer. When the system is right for months, people start checking less. Then the first serious error goes unnoticed. Not using AI is no longer an option. Without AI, you risk falling victim to fraudsters who use it, or losing to competitors whose processes run better with it,” Azizov says.

In his view, the main protective measure is understanding of the processes and reasonable limits on AI actions. “You can only delegate a process you understand really well. Without that, you won’t notice when the AI goes wrong. High-risk decisions must stay with humans.”

Deya Innab, Eastnets: Knowing the customer is no longer enough

Deya Innab, deputy CEO of compliance and payments technology company Eastnets, warns that identifying the account holder no longer answers every security question when an AI agent is making decisions. In a July 2026 article for the World Economic Forum, she explained how false information or fraudulent requests could manipulate an agent into acting through an otherwise legitimate account. “An agent that can act quickly across several systems can also make mistakes quickly,” she wrote.

Her proposed measures focus on traceability: institutions should record when an agent was involved and be able to explain why a suspicious payment was approved or a legitimate one blocked.

Sebastian Siemiatkowski, Klarna: Cheaper service can hide a quality problem

Sebastian Siemiatkowski, co-founder and CEO of Klarna, offers a warning grounded in his company’s experience. In May 2025, he acknowledged that emphasizing cost in AI-assisted customer service had led to lower quality.

Klarna had previously reported that its assistant performed work equivalent to 700 customer service agents. The company subsequently began piloting additional human support, with a spokesperson confirming that customers should always have the option to speak to a person.

Aleks Mayflower, Moonbot.pro: Human expertise can be lost

Aleks Mayflower, COO of Moonbot.pro, warns that AI’s biggest blind spot may be the expertise people stop developing when machines do the work for them. “One of the biggest risks of AI is that we may lose the hands-on experience that teaches us to question conclusions and make our own decisions,” he says.

Understanding why a system acts is as important as knowing the risks. While a convincing answer is easy to obtain, judging its assumptions takes practical experience, and even intuition reflects lessons accumulated over time.

That creates a problem for the next generation. Experienced professionals can challenge automated decisions today, but newcomers need opportunities to develop the same judgment. If AI takes over the tasks through which people once learned, that foundation may weaken.

“As we delegate execution, we risk gradually handing over something far more consequential: the decisions themselves,” Mayflower says.