OpenAI model "out of control" sounds the alarm, Microsoft quickly launches low-cost cybersecurity model

According to Zhitong Finance APP, amid the heated discussion surrounding the "out of control" incident involving OpenAI's model, Microsoft (MSFT.US) launched a new artificial intelligence (AI) model for discovering cybersecurity vulnerabilities on Monday. This is the first major product launched to revitalize its cybersecurity business since Microsoft reappointed former Google executive Hite Galot to lead the security department.

Microsoft stated that when used in conjunction with OpenAI's general large model GPT-5.4, its MAI-Cyber-1-Flash model outperformed Anthropic's Mythos 5, Google's 3.5 Flash Cyber, and OpenAI's GPT-5.5 Cyber in the CyberGym benchmark test.

Mustafa Suleyman, CEO of Microsoft's AI division, stated at a company event in San Francisco: "Our model's performance is world-leading, and the cost is only half that of competitors."

This generative model is Microsoft's first model aimed at cybersecurity scenarios. Galot introduced in a blog post that the model will be integrated into Microsoft's Project Perception. This project integrates multiple AI agents that can identify and fix system security vulnerabilities and will be open for public preview starting August 3.

She returned to Microsoft in February as Executive Vice President of Security Business, while Charlie Bell, the former executive responsible for this business from Amazon Web Services, transitioned to an independent technology researcher.

Generative AI has significantly lowered the threshold for hackers to exploit the latest publicly disclosed vulnerabilities to launch attacks. To this end, both Anthropic and OpenAI have launched defensive AI models aimed at cybersecurity professionals.

So far this year, Microsoft's stock price has fallen by 19%. A UBS analyst team led by Karl Keirstead wrote in a report on Sunday: "Given that the market generally believes open-source AI models will capture market share from frontier labs, investors are now reassessing Microsoft's deep ties with OpenAI and viewing it as a potential risk." However, Keirstead maintained a "buy" rating for Microsoft.

This year, Microsoft has successively launched several self-developed large models: first, a model that generates code in the GitHub Copilot tool, and recently, a native AI model integrated into Excel spreadsheet software. Microsoft CEO Satya Nadella maintains a collaborative relationship with OpenAI, but he is also investing computational resources to train self-developed models to improve operational efficiency.

Nadella posted on social media platform X on Monday: "By combining dedicated models, industry data with adaptive agents, tools, secure environments, and systems, we can enhance cost-effectiveness."

Since 2023, Microsoft has not disclosed the revenue scale of its cybersecurity business, which surpassed $20 billion in annual revenue that year.

In 2023, Microsoft launched the "Security Copilot" for cybersecurity professionals, which integrates OpenAI's GPT-4. This service is now pre-installed in two of Microsoft's high-end office software suites. Project Perception, once authorized, can suggest and implement code changes and can connect with non-Microsoft products Galot stated in an interview that executives in the cybersecurity industry view this tool as a key to breaking the deadlock: it can lower the entry barrier for security operations talent, allowing Security Operations Centers (SOC) to attract more personnel, as such talent is currently very limited in the industry. Companies typically have Security Operations Centers (SOC) equipped with a large number of personnel responsible for monitoring threats to information technology systems.

OpenAI model out-of-control attack exposes industry risks, AI safety barriers receive significant attention

Last week, OpenAI stated that its model exploited vulnerabilities to attack the infrastructure of AI startup Hugging Face. This incident exposed the potential risks of cutting-edge large models under autonomous decision-making, triggering strong industry attention on AI safety barriers.

Bipartisan lawmakers in the U.S. Congress proposed the AI Kill Switch Act, aimed at granting federal agencies the authority to suspend the operation of AI models when necessary. It is currently still in the proposal stage in Congress and has not officially taken effect.

Additionally, reports indicate that OpenAI CEO Sam Altman and NVIDIA CEO Jensen Huang will meet this week with Senate Intelligence Committee Democratic Chairman Mark Warner to discuss AI development and safety issues.

Microsoft AI head Suleiman described the OpenAI cyberattack incident as a "wake-up call" in the cybersecurity field. Suleiman pointed out that this is an important signal of the rise of AI cyberattacks. "As models become more powerful, the principle of prevention becomes very important."

Galot also commented, "This case fully demonstrates that in the face of cyber attackers wielding AI tools, we must rely on AI technology to build a defense system."

Microsoft's new cybersecurity model still has ample room for performance optimization.

Suleiman revealed in the interview, "We hold a unique massive security dataset, and currently, less than 1% of the data reserve has been used to train the model."