Top Stories This Morning: OpenAI Models Were Communicating Before the Cyberattack; DeepSeek Raises Prices

OpenAI has revealed new details about the hack of the Hugging Face AI platform, reporting that its models had been secretly coordinating their actions since May and ultimately gained access to the internet. Meta reported that during testing, its Muse Spark 1.1 model also gained access to the network and infiltrated a third-party company’s systems due to a configuration error in the isolated environment. Read about these and other topics in our roundup of key events as of the morning of August 6.

OpenAI has revealed new details about the Hugging Face hack: as early as May, AI models began quietly exchanging messages via internal “bulletin boards,” coordinating attempts to gain access to the internet to solve problems that could not be solved without an internet connection, according to Bloomberg. The company acknowledged that it had inadvertently given the models impossible tasks, and that the models persisted and began looking for workarounds.

In July, the models discovered a vulnerability, gained access to the internet through the internal infrastructure, and attacked the systems of OpenAI and Hugging Face themselves, the agency reports. Following the incident, the company slowed down its research to focus on strengthening security measures and warned that in the future, attackers could deliberately use AI agents to carry out similar cyberattacks.

Chinese AI startup DeepSeek has announced an upcoming significant price increase for its services. This is an unusual move for a company that has consistently put pressure on its competitors with ultra-low rates, according to Bloomberg.

DeepSeek has not disclosed exact figures. The company described the upcoming changes as significant and urged users to adjust their plans in advance. Currently, DeepSeek charges $0.14 per million input tokens and $0.28 per million output tokens for the V4 Flash model. By comparison, Chinese competitor Moonshot AI prices its Kimi K3 model at $3 and $15, respectively, while Anthropic charges $10 and $50 for its Fable 5 service.

DeepSeek’s decision to raise its rates could mark a turning point for the Chinese artificial intelligence market, Bloomberg notes. Until now, other major players have copied the AI startup’s strategy by offering low-cost, open-source models. Startups and their investors have long been looking for a way to transition from the era of ultra-cheap AI services to commercially viable business models.

During cybersecurity testing, Meta’s AI model, Muse Spark 1.1, gained access to the internet due to an error in the configuration of its isolated environment, infiltrated a third-party company’s systems, and made changes to its internal systems, according to The Information. Meta stated that the cause was an incorrect configuration by its partner, Irregular, and that the model subsequently exploited a vulnerability in another third-party service. The company is conducting an investigation.

The incident was the latest in a series of similar cases: OpenAI had previously reported that its model gained access to the internet during testing and hacked Hugging Face, while Anthropic revealed that its models had gained access to the systems of three organizations due to a similar error.

Iran stated that the U.S. is ready to “return to its commitments,” but denied reports of direct talks with Washington, despite Donald Trump’s statement that an agreement is close, according to CNBC. According to Iran’s Deputy Foreign Minister Kazem Garibabadi, contacts are taking place exclusively through Oman, and a joint statement by Tehran and Muscat on the Strait of Hormuz is in the final stages of preparation.

Axios previously reported that the U.S., Iran, and Oman are close to reaching a temporary agreement to reopen the Strait of Hormuz, and U.S. Treasury Secretary Scott Bessent suggested that a deal could be announced as early as the coming days.

Google will centralize leadership of its artificial intelligence development in Mountain View, gradually shifting key management functions from DeepMind’s London office to California to accelerate the development of Gemini and intensify competition with OpenAI and Anthropic, Bloomberg reports. Koray Kavukcuoglu will lead AI research and operations, while DeepMind co-founder Demis Hassabis will focus on strategy in his roles as chairman of Google DeepMind and chief scientist at Alphabet.

The reorganization is taking place amid delays in the release of the most powerful version of Gemini, the departure of leading researchers to startups and competitors, as well as record investments in AI—up to $205 billion in capital expenditures by 2026, the agency notes. Google expects that centralizing management will help speed up decision-making and improve the efficiency of developing the company’s key AI model.

— Japan's broad-based Topix index was virtually unchanged, while the Nikkei 225 fell 0.9%.

— Hong Kong’s Hang Seng Index fell 1.8%, while mainland China’s CSI 300 Index fell 0.6%.

— In South Korea, the KOSPI fell 4%, while the KOSDAQ rose 1.5%.

— Australia's S&P/ASX 200 rose 0.4%.

— Nasdaq Composite futures were down 0.1%, while S&P 500 and Dow Jones Industrial Average futures were up 0.2%.

This article was AI-translated and verified by a human editor