Open-Source AI Models Like DeepSeek Exploited by Cybercriminals, Triggering Surge in Bank System Vulnerability Scanning

Overview

Recently, open-source or low-cost AI models such as DeepSeek and Kimi (by Moonshot AI) have been widely abused by cybercrime syndicates, leading to a surge in requests for vulnerability detection against financial institution systems. Compared to traditional closed-source models, these models are cheap, have publicly available weights, and lack effective behavioral tracking, making them ideal tools for attackers conducting vulnerability mining and penetration testing, posing severe security challenges to banks and other critical infrastructure.

Security Risks of Open-Source AI Models

Unlike closed-source models such as Anthropic's Claude, open-source models like DeepSeek and Kimi pose significant security risks. Closed-source models typically run in cloud environments controlled by their developer companies, where any anomalous usage is monitored in real time, accounts can be frozen rapidly, and user logs can fully trace identities and corporate information, resulting in high crime costs and high exposure risks.

However, open-source weight models offer a completely different attack path:

• Massive cost advantage: API call fees are only one-tenth or even one-hundredth of those for closed-source models, allowing attackers to invoke them indefinitely without worrying about economic burden.
• Security filters can be stripped: Using specialized tools, attackers can forcibly remove the model's ethical constraints and safety filtering mechanisms and deploy it on their own servers.
• Fully anonymous operations: In self-hosted modes, attackers do not need to expose their IP addresses, identities, or corporate information, achieving a true "invisible" attack environment.

"Cybercrime syndicates are using open-source AI for automated vulnerability mining, forming a large-scale, low-cost gray industry chain."

Banking Defense Systems Under Pressure

According to reports from Nikkei, several major domestic banks discovered this summer that the number of system vulnerabilities detected with the aid of cutting-edge AI doubled compared to the same period last year. This phenomenon directly reflects the rapidly rising prevalence of AI-driven cyberattacks.

Facing increasingly complex attack methods, banks are adopting a dual-response strategy:

1. Accelerating vulnerability patching: Immediately implementing code fixes and architecture optimizations for high-risk defects already detected.
2. Reserving contingency plans: If repairs cannot be completed within the attack window, some banks have included "system autonomous shutdown" in their emergency response plans to prevent data breaches or malicious exploitation of services.

This disclosure marks the first detailed media report on the banking sector's internal response since Anthropic released advanced models such as Claude 3.5 Sonnet, signaling that the AI security topic has moved from the technical circle into regulatory and financial operational practice.

Industry Reflection and Outlook

This incident has triggered in-depth discussions within the industry regarding the governance boundaries of open-source models. On one hand, open-source technology has promoted AI democratization and innovation; on the other hand, an uncontrolled open environment can also be reverse-engineered and exploited by malicious forces.

In the future, model providers may need to strengthen:

• Server-side behavioral auditing and anomalous traffic detection
• Dynamic key management and tiered usage permission mechanisms
• Threat intelligence sharing channels with financial institutions

Meanwhile, regulators also need to accelerate the formulation of laws and regulations addressing the abuse of AI tools, balancing the relationship between technological innovation and cybersecurity.