The Zhitong Finance App learned that Open Source Securities released a research report saying that with the accelerated implementation of AI application scenarios such as large models and smart devices, security risks brought by AI penetrated into core business processes, and also spawned new systemic risks such as reminder injection, model poisoning, and multi-agent collaborative deception. In 2027, AI application security will remain the largest category of expenditure, which is expected to reach nearly US$851 million, followed by AI usage control, which is expected to reach US$749 million. In the AI era, cybersecurity has entered the stage of “AI versus AI”, and the industry has also ushered in new development opportunities.
The main views of Open Source Securities are as follows:
In the AI era, cybersecurity has entered the stage of “AI against AI”
With the widespread application of artificial intelligence technology, the field of cyber attacks has ushered in a critical iteration. Attackers have begun to use AI to complete target screening, information integration, path planning, and implementation of attacks and strategy adjustments, and use AIAgent to launch automated attacks, so that the rules of the cybersecurity attack and defense game have evolved from people to the stage of “AI versus AI” in the past. At the same time, as AI applications such as big models and smart devices accelerate implementation in government, finance, industry, medical care, and internal enterprise management scenarios, AI brings security risks that penetrate core business processes, and also spawn new systemic risks such as reminder injection, model poisoning, and multi-agent collaborative deception.
Anthropic and OpenAI launch AI security defense models, attach great importance to AI security
(1) According to Anthropic's disclosure, a hacker group tracked down as GTG-1002 has launched a highly automated infiltration operation against global targets using Claude Code tools. Instead of using any self-developed malware, the attackers combined publicly available penetration testing tools to build a multi-agent attack architecture using Claude Code and MCP (Model ContextProtocol). They convinced Claude that he was an employee of a legitimate cybersecurity company and was doing authorized security tests. Since then, Anthropic has launched the Claude Mythos project—an AI system dedicated to security defense. It discovered thousands of zero-day vulnerabilities during testing in April 2026, covering all major operating systems and browsers. In addition, OpenAI launched GPT-5.4-Cyber as a specially optimized version of GPT-5.4, which is specially optimized for cybersecurity defense scenarios. (2) On August 27, OpenAI released an open letter entitled “Collective Action Calls for Cyber Defense”, which was co-signed by at least 116 companies and organizations, including Anthropic, Google, Microsoft, AMD, CrowdStrike, Palo Alto Networks, Okta, Fortinet, Citibank, Capital One, Visa, Mastercard, and other heavyweight institutions spanning the fields of AI, cybersecurity, finance, and infrastructure. The central warning of the open letter is that AI-driven cyber attacks will become “more widespread and sophisticated” in the coming months.
As the AI penetration rate increases, the global AI security market will achieve rapid growth
Gartner predicts that in 2027, the global AI security market will grow 69% from 2026 to US$4.783 billion, and further rise to approximately US$7.7 billion in 2028. The rapid growth of the AI security market is mainly due to the urgent need for enterprises to protect artificial intelligence systems, resolve emerging vulnerabilities, and strengthen defense against complex cyber threats. In 2027, AI application security will remain the largest category of expenditure, which is expected to reach nearly US$851 million, followed by AI usage control, which is expected to reach US$749 million.
Risk warning: Technological progress falls short of expectations, and market competition increases risk.