Charlie Bell, Microsoft Executive Vice President for Security, introduces new security products in San Francisco, U.S., on Sept. 27 (local time). Correspondent Kim Chang-young
As hacking powered by artificial intelligence (AI) grows more serious, AI tools designed to defend against AI attacks are drawing attention.
Microsoft (MS) on the 27th unveiled an agent-based security platform called “Project Perception” and its first proprietary security-dedicated AI agent, “MAI-Cyber-1 Flash,” in San Francisco.
Microsoft said that running MAI-Cyber-1 Flash on Project Perception can deliver industry-leading model performance at 50% of the cost. When the model was used together with the general-purpose GPT-5.4, it recorded a CyberGym score of 95.95% in benchmarks. That figure is more than 10 percentage points higher than OpenAI’s GPT-5.5 Cyber (85.6%) or Anthropic’s Mithos5 (83.8%).
Microsoft stressed that defensive AI has become important as AI hacking spreads. Hyatt Gallott, Executive Vice President (EVP) of Microsoft’s security division, introduced news that 90% of last year’s cyberattacks by Chinese hackers were AI-driven crimes, noting, “Humans merely gave instructions to the AI agents, and the agents processed 15,000 requests per second. That is a scale and speed humans cannot handle.” She added, “The situation has now completely changed. With AI, attackers are gaining an edge over defenders,” introducing Perception as “an agent-based security system for defenders, designed to defend AI with AI.”
Following the emergence of Anthropic’s “Mithos,” an incident in which OpenAI’s GPT model performed unauthorized hacking of an AI platform has heightened security concerns. OpenAI, Anthropic and Google, which are competing to advance AI models, are warning of the risks of AI models while simultaneously releasing security tools.
Mustafa Suleyman, CEO of Microsoft’s AI division, said in an interview with the Financial Times (FT) regarding the OpenAI incident, “These are very powerful tools, so they must be handled with extreme care.” He noted, “As models become more powerful, the principle of prevention will become important. This incident is a kind of warning.”
Mustafa Suleyman (right), CEO of Microsoft AI, introduces “MAI-Cyber-1 Flash” in San Francisco, U.S., on Sept. 27 (local time). Correspondent Kim Chang-young