Global artificial intelligence (AI) corporations such as Anthropic, OpenAI and Microsoft (MS) are unveiling security AI systems, intensifying competition in security systems. As AI corporations expand their businesses from Generative AI to security systems, some note that the security industry is being reshaped.
According to the security industry on the 5th, Anthropic, the developer of the Generative AI model Claude, on the 2nd (local time) added about 150 institutions in 15 countries to the participants in Project Glasswing, a cybersecurity consortium that restricted access to Mythos. In Korea, Samsung Electronics, SK hynix, SK Telecom and Korea Internet & Security Agency (KISA) were included.
Mythos, which Anthropic unveiled in April, is a high-performance security AI system specialized for detecting security vulnerabilities. It is assessed to be capable of having AI design its own intrusion paths and verify the feasibility of attacks. It shocked the industry by showing the potential to process vulnerability analysis tasks—previously performed by security experts over several days to weeks—within minutes to hours.
Initially, Anthropic granted access to Mythos only to about 50 corporations including Apple and Google. The company was concerned that Mythos's advanced coding capabilities could be misused for hacking. However, as the participant pool expanded this time, Korean corporations such as Samsung Electronics and SK hynix, which recently made billion-dollar-level investments in Anthropic, joined.
Since Anthropic unveiled Mythos, global AI corporations have been rolling out security systems in succession. OpenAI unveiled GPT-5.4 Cyber in April and then introduced GPT-5.5 Cyber to selected partners in early May. At the same time, it launched the Trusted Access in Cyber (TAC) program, a cybersecurity cooperation initiative.
Unlike Anthropic, OpenAI is actively working to expand the security ecosystem. OpenAI is accepting applications to join TAC through its official website and recently expanded the list of participating countries in Daybreak, a global security initiative for national and public institutions. Jason Kwon, OpenAI chief strategy officer (CSO), said at a recent press briefing in Korea that Korea and Japan will newly join Daybreak.
MS also unveiled in May its own security AI system, the Multi-model Agentic Scanning Harness (MDASH). Although MS entered the race as the third mover compared with competitors, it moved to catch up by adopting a multi-agent architecture instead of the single-model approaches used by rivals. More than 100 specialized AI agents are designed to perform the entire process end to end, from vulnerability discovery to verification and debate to proof of exploitability.
MDASH topped the public CyberGym benchmark leaderboard with a success rate of 88.45%, the highest score. That is about 5 percentage points (p) higher than Anthropic's undisclosed frontier model Claude Mythos Preview (83.1%), which ranked second.
This competition shows that the security industry is being reshaped. In the past, security corporations competed around threat detection and response technologies, but now analysis indicates that global AI corporations with large-scale AI models and cloud infrastructure are expanding their influence into the security domain. Corporations developing AI models are simultaneously evolving into security platform providers. In this process, corporations' participation in AI security ecosystems is also emerging as a task.
A security industry official said, "More security models applying AI will continue to emerge," and added, "As the performance of AI security models advances rapidly, it is now important for corporations not only to introduce individual security solutions but also to decide which security AI ecosystem to join."