
Yoido Full Gospel Church in Seoul said personal data tied to 850,000 members may have been stolen, with KISA flagging the breach and Oasis Security finding the data on an overseas server alongside attack logs suggesting AI sub-agent use; Sarang Church also reported a smaller breach affecting 89,000 members and 286 employees.
AI-generated summary
Yoido Full Gospel Church is a Seoul megachurch once recognized by Guinness as the world's largest congregation. South Korea has seen increased cyberattacks, including on financial institutions, with officials noting AI's potential role in recent breaches.
Yoido Full Gospel Church, the Seoul megachurch Guinness once recognized as the largest congregation in the world, said Wednesday that personal data tied to 850,000 members may have been stolen.
The church's initial review found names and dates of birth, plus a log of changes members had made to their records. "The file contained 2,629 changes to resident registration numbers, 3,964 changes to phone numbers and 7,202 changes to addresses," the church said in a statement.
South Korea's internet security agency, KISA, flagged the suspected breach to the church on Tuesday afternoon. The church has since blocked outside access to its systems, changed its server passwords and begun notifying members.
Oasis Security, a cybersecurity company, found the data on an overseas server that also held attack logs and account details linked to Yoido and a second Seoul megachurch, Sarang Church. The firm came across it in September while tracing internet addresses tied to suspected attacks.
Sarang's haul is smaller: about 89,000 member records and 286 employee records, including the senior pastor's. The logs suggest that data was stolen in August, and the church has formed an emergency task force and reported the incident to authorities.
Oasis said the attack records show activity of AI sub-agents, meaning helper programs that an AI model launches to handle separate chunks of a job, along with long attack reports that look machine-written. The churches are still working out how the intruders got in, and AI's role remains unclear.
President Lee Jae Myung confirmed Tuesday that AI is believed to have been used in recent hacks on South Korean commercial banks. Shinhan Bank's breach exposed names, phone numbers, annual income and borrowing limits for about 25,000 customers, according to local news reports, and the financial regulator opened an emergency on-site inspection.
Yoido Full Gospel Church said it plans to replace its firewall, the filter that blocks unwanted traffic into a network, and bring in security firms to hunt for other weaknesses while it notifies the 850,000 members whose data may be affected.
AI outlook — possibilities, not facts
Yoido Full Gospel Church will complete firewall replacement and security audits within the next few months
Likely · Within months
South Korean authorities will increase scrutiny of AI's role in cyberattacks following these incidents
Likely · Within weeks

OpenAI published 722 math manuscripts on GitHub generated by an unreleased internal model from a single prompt, with only 22% formally verified in Lean; experts debate the significance, reproducibility, and responsibility of AI-produced mathematical results.

Anthropic released Claude Haiku 5.5, positioning it as the cheapest, fastest, and most capable small model for high-volume tasks like document summarization and live customer support. Priced at $0.10 per million input and $0.50 per million output tokens, it offers up to 75% average savings over Haiku 4.5 and matches OpenAI's GPT-6 Luna pricing. Benchmarks show Haiku 5.5 outperforms Luna on OSWorld 2.1 (72.4% vs 48.9%) and Terminal-Bench 4.0 (39.2% vs 16.4%), while scoring 1620 on GDPval-AA v2.1. The model includes an adjustable effort setting and follows Opus and Sonnet 5.5 releases. Available on Claude website, AWS, Google Cloud, and Azure as claude-haiku-5-5, with new monthly API credits for Max and Team plans.

Google launched Playground, a web platform allowing users to create playable games by describing them in plain language, with refinement via chat and sharing via link. Available to U.S. users 18+, access tiers depend on Google AI subscription. Games run in browsers, support privacy or community publishing with safety screening, and may integrate Unity Spark for advanced features. The launch follows Google's Genie AI projects and coincided with a drop in game stocks as investors worry about AI disrupting traditional game development tools.

Google launched Nano Banana 2.1, an updated AI model that generates and edits images from text, now available in Gemini app, Search AI Mode, Ads and developer tools. The model offers better visual design, mask-based editing and subject consistency, scores 1,050 ELO in preference tests, and reduces API costs to $0.0336 per 1K image, half the price of its predecessor.

Developer hashfunction published metadata for 5.6 billion public TikTok videos spanning 2014 to 2026 on Hugging Face and datasocial.ai. The 460 GB dataset includes captions, engagement metrics, and AI flags, collected via a reverse-engineered mobile API.

Mistral AI launched Mistral Large 4, a 1-trillion-parameter AI model using a mixture-of-experts design with 49 billion active parameters per query. The model is positioned as a competitive open-weight alternative to Claude Opus 5.5 and GPT-6 Astra, with pricing at $1.36 per million input tokens and $4.18 per million output tokens. Benchmarks show strong performance in coding and automation tasks, trailing only Kimi K3 and Gemini 4 Argon in some tests. Mistral plans to release the model weights by end of October.