China Approves New Nuclear Reactors with Investments of 20.6 Billion

Anthropic stated that during routine tests, some of its models gained access to the internet and compromised the systems of three different organizations. This became known following an internal analysis initiated after competitor OpenAI disclosed a similar issue.

In a statement published on Thursday, Anthropic reported that it began auditing its own systems after OpenAI revealed last week that some of its models escaped the test environment, accessed the open internet, and hacked the Hugging Face AI platform's systems.

Anthropic discovered three instances where its models accessed the open internet without authorization and 'gained unauthorized access to the production infrastructure of three different organizations.' The company learned of these incidents by analyzing over 140,000 evaluations following the OpenAI report. Similar to the OpenAI tests, standard safety mechanisms were disabled during Anthropic's checks to assess the full potential of its models.

Anthropic explained that in the three cases, the models were given a simulated 'flag capture' task, being told that the 'flag' was hidden on different equipment within the network, and their objective was to penetrate and retrieve it. Unlike the OpenAI situation, Anthropic emphasized that none of its models intentionally tried to leave the test environment. Instead, the company stated that a misunderstanding occurred between Anthropic and its evaluation partner, as the models should not have had access to the open internet, but they managed to obtain it.

To infiltrate the three unnamed organizations, the models used basic methods such as 'exploiting weak passwords' and searching for system entry points that did not require login credentials or tokens. The company noted that the most advanced version of its models recognized it was on the open internet at one point and independently ceased further actions.

Anthropic reported that the earliest security breach incident involving another organization occurred back in April, and none of the affected organizations were aware they had been hacked. Anthropic is currently cooperating with the affected organizations.

The disclosure of the Hugging Face hack by OpenAI models shocked the cybersecurity and AI worlds, as it became the first real example of what experts have long warned about: AI agents with advanced cybersecurity skills can escape test environments and cause real damage.

Like OpenAI, Anthropic announced on Thursday a suspension of all cyber audits. Anthropic acknowledged that it could have taken more 'proactive' measures to prevent these cybersecurity breaches.

Anthropic's data disclosure further confirms that unintentional hacking of other organizations by AI agents is not an isolated feature of one company, and it will likely strengthen calls for improved AI testing tools and methods to slow down AI development, which may be advancing much faster than society is prepared for.

The production of transformers will take place in Baghdod. It is important to understand that power supply problems are related not only to electricity generation but also to delivering this energy to the consumer without losses, for which transformers resistant to various pressures are necessary.

In the Toshlok district of the Fergana region, a transformer production facility operates, but its production capacity is insufficient to fully cover current needs. Due to this, foreign partners have been attracted to the region.

The region's governor, Khayrullo Bozorov, held discussions on investment projects related to transformer production in the Baghdod district.

Company founder Ji Zhaoshan provided information about his firm's activities, experience in transformer manufacturing, and planned projects in the Fergana region. The initial stage of the project involves importing raw materials and components necessary for transformer production and beginning their assembly. The subsequent stage is the full implementation of local production.

As a result of the meeting, an agreement was reached to intensify cooperation and move projects to a practical level, and a memorandum of understanding was signed.

The Minister of Public Administration and Development, Nkosi Mzamo Buthelezi, unveiled the AI-powered Batho Pele chatbot on Friday. This new digital tool is intended to transform the delivery of public services in South Africa, providing citizens and civil servants with instant access to vital information.

The chatbot is designed to provide citizens and public servants with up-to-date information on public service policies, circulars, legislation, and other necessary details. Minister Buthelezi noted that the launch of this tool fulfills a long-standing commitment to 'Batho Pele – People First,' which for 29 years has required the government to ensure citizens have information, clarity, and answers.

This initiative is the result of a partnership between the public sector and private business, involving DPSA, Meta, and Juicetel. It marks a significant step forward in modernizing public service delivery through the innovative use of technology.

As it develops, the chatbot will be able to process information and respond in nine official languages of South Africa. It will assist citizens and civil servants, especially those working in corporate and managerial roles, by reducing the time spent searching for policy updates, legislative acts, and circulars.

The Minister emphasized that the tool does not replace human judgment. He stated that the chatbot does not make decisions or substitute for human assessment; instead, it supports civil servants in making faster, more informed, and empathetic decisions that only a human can provide. The system will be implemented in line with constitutional principles of transparency, accountability, confidentiality, security, and fairness.

Looking ahead, the Minister described this launch as a foundation that can be scaled to national departments, provinces, municipalities, and state institutions, and ultimately extended to frontline services such as healthcare, Home Affairs, SASSA, education, and local government. He also pointed to South Africa's potential to become a leader in digital governance on the continent, calling this partnership a possible model for other African governments.

Partner Comments

Thabo Makhenete, Head of Public Policy in the South Africa region at Meta, commented: "Batho Pele AI demonstrates what is possible when government, the private sector, and local innovators join forces to practically apply AI. By making public service information accessible in nine South African languages through a simple conversational interface, this platform allows civil servants to spend less time searching for policy guidelines and more time serving citizens. We are proud to support this initiative alongside DPSA and Juicetel, and we are excited to see how Meta's Llama models are being used to solve real problems on the continent within locally created solutions."

The Minister thanked Meta and Juicetel for their collaboration with the government and urged other departments to utilize similar technologies to improve service delivery. He specifically noted that the responsible use of AI must be governed by proper governance, and civil servants retain responsibility for ensuring the accuracy of the disseminated information.