24H Flash Updates

According to a Fortune interview, OpenAI CEO Sam Altman confirmed the company will not pursue an IPO by 2026, stating that going public at this time is "ill-timed" given current AI safety concerns. Altman emphasized that OpenAI is not in a rush to go public, and will consider an IPO only when its business is ready and society is prepared to adapt to AI models of varying capability levels; the company must prioritize addressing AI safety and alignment issues, and may even slow down frontier model development once AI capabilities reach a new stage.

Earlier reports indicated OpenAI had considered delaying its IPO until 2027, with a potential valuation reaching $1 trillion. Meanwhile, Anthropic CEO Dario Amodei recently called for slowing down the advancement of frontier AI and proposed granting independent evaluation bodies access at the level of internal employees. Altman responded by expressing agreement, pledging that independent assessors will also be granted similar employee-level access, and stated more details will be shared shortly.

According to reports, Anthropic has announced a long-term on-site presence for third-party security teams, providing them with tools and access nearly equivalent to those of internal risk assessment teams. External reviewers can inspect training pipelines, safety protocols, and incidents, and independently publish their findings; Anthropic may only remove information related to security, legal compliance, or customer privacy, but cannot delete content due to unfavorable conclusions. CEO Dario stated that this approach goes further than any current AI company.

He also shifted his stance on "slowing down," noting that AI is increasingly involved in the development of next-generation AI systems, accelerating capability advancements—making pure safety research insufficient. The model's own capabilities must also be slowed down. He cited OpenAI’s Hugging Face incident as a warning, cautioning that within 6 to 12 months, similar agent swarms could establish botnets spanning large-scale internet infrastructures, potentially causing trillions in losses. He urged leading AI companies in democratic nations like the U.S. to unify safety standards and R&D speed, then attempt global coordination with China on AI development pace, while continuing to restrict advanced chip access and model distillation to maintain U.S. technological leadership.

It was reported that the UK-based fintech company Revolut confirmed a customer data breach, where an unauthorized third party exploited legitimate government agency domain email addresses to submit fraudulent information requests, resulting in the exposure of sensitive customer information. The leaked data includes identity and contact details such as date of birth, postal and email addresses, phone numbers, as well as copies of identity documents like passports and driver’s licenses; verification selfies, account statements, and transaction histories may also have been compromised.

Revolut stated the number of affected customers is "limited," and has directly notified relevant customers, though it did not disclose the exact figure or specify whether the incident was confined to particular markets, nor would it reveal the involved government agency. The company said it has blocked the associated email addresses and informed government bodies, law enforcement, and financial regulators, emphasizing that its systems and customer funds remain unaffected. Revolut serves over 80 million customers globally and operates banking services in more than 30 countries. Cryptocurrency security researcher ZachXBT noted that the incident appears to target high-net-worth individuals.

Reports indicate that over the past year, a new class of alerts triggered by AI tools and agents has emerged within enterprise Security Operations Centers (SOCs), growing at a pace exceeding all other alert types. These alerts are not indicative of attacks targeting AI itself, but rather reflect routine footprints of AI usage across organizations, including developers running coding agents and non-technical personnel integrating consumer-grade AI tools into enterprise environments.

Details of a coordinated attack targeting RubyGems, disclosed by Maciej Mensfeld, Senior Product Manager for Software Supply Chain Security at Mend.io, reveal that this "major malicious campaign" was carried out by a group of OpenAI agents, enabling remote code execution (RCE) on RubyDoc servers. The findings are based on a new report released by Spencer Kitts, Thomas Larsen, and Sydney Von Arx.

According to reports, James van Geelen, founder of Citrini Research—the organization behind the "AI Apocalypse Report"—has sold the company to SemiAnalysis, a semiconductor and AI research institution, though the transaction amount remains undisclosed. Van Geelen will temporarily continue serving as CEO of Citrini and plans to launch a new fund. Citrini currently boasts approximately 260,000 subscribers on Substack. The firm gained prominence in February this year after releasing the "Global Intelligence Crisis 2028" report, which envisioned AI rapidly displacing white-collar jobs, sparking widespread market discussion on AI's impact on employment and the economy, and briefly causing noticeable declines in software equities.

SemiAnalysis was founded by Dylan Patel in 2020 and has long focused on semiconductor and AI infrastructure research. Patel stated that this acquisition aims to scale up research operations and explore new models for investment research in the AI era. Van Geelen did not disclose specific details about the new fund.

Today, according to an official announcement, Tibo, Product Lead for OpenAI Codex, issued an update regarding recent quality issues reported by Astra users, stating that the team has identified and fixed multiple defects and will perform a reset before midnight local time today (12:00 PM Beijing Time).

Fixes include: abnormal behavior in legacy model skill triggering, disabling an optional context management experiment (affecting approximately 4,000 to 5,000 users), and removing improperly configured engines. Additionally, the team has implemented minor improvements, resulting in more consistent execution, improved accuracy in tracking users' latest messages, and enhanced result validation.

According to two informed sources, Anthropic is in talks with Nvidia to bring the latter on as an anchor investor for its upcoming IPO. The sources said Anthropic aims to raise up to $100 billion, potentially valuing the company at around $200 billion;

Nvidia is considering investing up to $10 billion in this IPO. The plan remains under discussion and could still change. The report noted that bringing in Nvidia could bolster investor confidence in the offering, which will serve as a major test of public market appetite for the massive valuations and capital demands of cutting-edge AI labs.

It is reported that Mecka AI, a robotics training data startup, is nearing the completion of a new funding round led by Sequoia Capital, with an estimated valuation of around $500 million. This round comes just three months after Mecka's previously announced $60 million Series A financing, though the exact size and final terms of the deal remain undisclosed. Founded in 2024, Mecka AI’s core business revolves around collecting and analyzing human motion data: users record first-person footage of everyday tasks—such as brewing coffee or repairing vehicles—via body sensors and smartphones, and Mecka AI compensates them for this data, which is then used to train general-purpose and humanoid robots.

This model bears resemblance to Scale AI’s data services for large language models. As of early June, Mecka anticipates achieving an annual run rate of $100 million by end-2026. Currently, robotics companies and AI labs widely rely on real-world data and teleoperation techniques to build their models. Similarly, another data-focused startup, XDOF, was previously reported to be close to closing a new funding round with a valuation reaching $1.2 billion.

According to reports, Y Combinator CEO Garry Tan publicly stated that U.S. regulators should not interfere with AI distillation practices and believes U.S. open-weight labs should also adopt distillation techniques to learn from cutting-edge closed-source models, forming an "American distillation mechanism" to provide more open-weight alternatives independent of China. Distillation refers to the process of training one model by leveraging extensive prompts to mimic another model's reasoning capabilities, a practice widely recognized as legitimate within the industry for model training.

Tan emphasized that he does not advocate for illicit distillation involving stolen credentials, but rather hopes U.S. labs can access models through legitimate "front-door" entry points. He also argues that leading labs should not restrict users' usage of model outputs, noting that closed-source models themselves were trained using vast amounts of human knowledge without explicit permission. In response to Anthropic's recent report alleging "illicit distillation attacks" by Chinese labs and its CEO Dario Amodei's call for regulatory crackdowns, Tan holds a differing view. He stated that the true AI doomsday scenario is when frontier AI capabilities are monopolized by a single proprietary entity.

According to reports, the Regional Finals of Central Eurasia’s Road to TechCrunch Startup Battlefield 2026 have been unveiled: Cerberus, WeGlobal AI, and LOOQ secured the top three spots, earning their representation of the Central Asian region at the Startup Battlefield 200 during TechCrunch Disrupt in San Francisco, October 13–15. This year’s competition received 726 applications from 39 countries, with 84% focused on AI-based products.

The three finalists will share a $100,000 investment pool funded by Astana Hub Ventures, IT Park Ventures, and Silkroad Innovation Hub: Cerberus receives $50,000, WeGlobal AI $30,000, and LOOQ $20,000. OpenAI provided a total of $280,000 in API credits to all 22 finalists. Cerberus (Uzbekistan) is an AI-powered cybersecurity startup automating the detection of vulnerabilities within enterprise IT systems; WeGlobal AI (Kazakhstan) leverages AI tools to monitor student mental health and provide career guidance in educational institutions; LOOQ (Japan) develops camera hardware and an “Attention API” to measure real-time viewership for outdoor advertising campaigns.

It was reported that UK-based AI data center startup Nscale has appointed Fidji Simo, a former executive at OpenAI, Meta, and Instacart, to its board. Simo previously served as CEO of OpenAI’s AGI Deployment division, effectively the second-in-command at the lab, and stepped down in July this year due to health reasons; she remains a part-time advisor to the company.

Previously, she led Instacart’s IPO in 2023 and spent over a decade at Meta. Nscale, founded just two years ago, specializes in the design, construction, and operation of AI data centers; according to Bloomberg's report last week, the company is seeking to raise up to $3.5 billion ahead of its planned IPO.

According to TechCrunch, Roblox announced the expansion of its generative AI-powered game creation capabilities at its annual Developer Conference (RDC). Its natural language-based creation tool "Build," initially launched in New Zealand in July, is now extended to Serbia and Singapore, with added desktop access, asset library integration, and iterative control features; later this year, it will introduce prompt-driven scene generation functionality and be integrated into Roblox Studio. Creators can now publish games as standalone applications across mobile, PC, and console platforms, while Roblox serves as the underlying game engine—players will be able to join games directly via browser links by year-end.

In terms of monetization, Roblox unveiled the Creator Wallet, partnering with Airwallex to offer daily settlement services to users aged 18 and above in the United States, and announced the upcoming launch of the Roblox Card. Platform data reveals that creators have collectively earned over $5 billion since 2013, with $1.7 billion generated in the past 12 months; in the last quarter alone, players across 180 countries accumulated over 29 billion hours of gameplay. The announcement underscores Roblox’s dual goal: attracting more users to the platform while expanding the distribution reach of developer content and lowering the barrier to entry for game creation.

According to an official announcement, Anthropic disclosed on Thursday that it had thwarted an activity launched by a Russia-backed threat actor. The actor abused Claude to develop AI-assisted workflows, attempting to reconstruct malicious software before detection. Anthropic attributed the incident to its tracked espionage group GTG-20006 (Generative Threat Group), which aligns with the previously reported Midnight cluster.