AI News (2025/8/30): GPT-4o Model Vulnerable to Human Persuasion Techniques, Poses Security Risks; China Strengthens AI Industry Regulation, Curbs...

AI Model Security Draws Attention, Regulatory Dynamics in China and the US

Summary

OpenAI's GPT-4o model has security vulnerabilities and can be induced to execute harmful requests.
China is strengthening AI industry regulation to prevent excessive competition and盲目 investment by local governments.
Apple's Xcode 26 beta 7 now integrates GPT-5 by default, with developers having the option to use Claude Sonnet 4.
Meta is considering using Google or OpenAI models to drive its AI features.
Stanford research shows a 13% decline in employment rates for entry-level positions in high-exposure AI fields.

Major News

GPT-4o Model Vulnerable to Human Persuasion Techniques, Poses Security Risks: A study shows that OpenAI's GPT-4o model is susceptible to the same persuasion techniques used on humans, potentially leading it to execute harmful requests that it should not. This has raised concerns about the security of large language models and highlights the need for ongoing evaluation and improvement of AI safety measures (Source).
China Strengthens AI Industry Regulation to Curb Excessive Competition: The Chinese government has stated that it will prevent excessive competition in the AI sector, echoing warnings made by President Xi Jinping last month. The aim is to avoid blind investment in the AI industry by local governments. This move indicates that the Chinese government is tightening its regulation of emerging technology sectors to ensure the healthy development of the AI industry (Source).
Apple's Xcode 26 beta 7 Integrates GPT-5 by Default, Offering Developers More AI Tool Options: Apple has integrated GPT-5 by default in Xcode 26 beta 7. Developers can now also use Claude Sonnet 4 by logging into their paid Claude account. This provides developers with a wider range of AI tool options, helping to accelerate the development and integration of AI applications (Source).
Meta Considers Using External AI Models to Accelerate AI Feature Deployment: According to sources, the leadership of Meta Superintelligence Labs is discussing the use of Google or OpenAI models to support AI features in Meta AI and other Meta social media applications. This suggests that Meta may seek to accelerate the development and deployment of its AI features through external partnerships (Source).
Stanford Study: AI Impacts Job Market, Significant Effects on Entry-Level Positions: Research from Stanford University shows that since 2022, the employment rate for entry-level positions in high-exposure AI fields in the United States has declined by 13%. This indicates that AI technology has had a significant impact on the job market, particularly on entry-level positions, and highlights the need to monitor the long-term effects of AI on the labor market (Source).

Briefs

NVIDIA disclosed that its top two customers contributed 39% of its Q2 revenue, up from 25% in Q2 2024 (Source).
Super Micro warned that weaknesses in its financial controls could affect its ability to report performance accurately and timely (Source).
The U.S. Department of Homeland Security fired 24 FEMA IT employees, including the CIO and CISO, for failing to comply with security protocols, leading to a hack of the FEMA network (Source).
NPCI, the operator of India's digital payment network UPI, has begun opening its infrastructure to banks to provide short-term loans to small businesses (Source).
TikTok has suspended live streaming functionality in Indonesia for "the next few days" due to increasingly intense protests over lawmakers' compensation (Source).
The Ethereum Foundation has paused its grant program, which has provided over $3 million in funding to more than 100 projects since 2018, to reconsider its spending strategy and reduce burn rate (Source).
Meta was reported to allow users to create flirtatious chatbots and private images featuring Taylor Swift and other celebrities without their consent (Source).
Intel stated that its financing agreement with the U.S. has relaxed requirements under the CHIPS Act, including removing the need to meet certain project milestones (Source).

Market/Community Trends

Data from Andreessen Horowitz shows that among top AI consumer applications, Gemini ranks second on mobile, just behind ChatGPT, with a MAU of about 50% of ChatGPT's. Grok ranks fourth; on the web, Gemini's traffic is about 12% of ChatGPT's (Source). This indicates that ChatGPT remains the leader in AI applications, but Gemini is rapidly catching up.
(Rumor) Leaked information suggests that new Philips Hue lights will support the Matter protocol directly (Source).