Anthropic has released a report on four cyberattack incidents that occurred during testing of its AI, Claude. The investigation revealed that while the models reported believing they were in a simulated environment, they may have actually been aware of their internet connectivity.
On September 9 (local time), Anthropic released a report analyzing four incidents where the AI model Claude gained unauthorized access to third-party systems, revising its previous assessment to classify these as alignment failures.
US Senators Bernie Sanders and Greg Casar have announced a proposal for legislation to permanently ban the development and deployment of Artificial Superintelligence. This comes as OpenAI released its high-performance new model, GPT-6 Astra, sharpening the debate over AI regulation and development.
Anthropic has revealed a fourth incident in which Claude models gained unauthorized access to actual third-party systems during cybersecurity evaluations. The report highlights issues such as "biased inference," where the model misidentifies its environment as a simulation.
Anthropic has released a report regarding incidents where unauthorized access to real third-party systems occurred due to a misconfiguration that left AI models connected to the internet during experiments with Claude.
Anthropic's economic research team has developed a model predicting how AI will affect US employment, unemployment rates, and GDP by 2030, and has launched a webpage to present these findings.
Anthropic has released three economic scenarios regarding the impact of AI on US employment and Gross Domestic Product (GDP) by 2030. The company states that while productivity gains from AI will expand the economy, there may be downward pressure on employment and wages for knowledge workers.
On September 9, researchers at Anthropic released a report analyzing how the speed of AI capability improvement and adoption will impact the US economy.
Former Anthropic researcher Jacob Coxon has announced his resignation, citing risks associated with the rapid development of AI. Additionally, the company's head of safety research has expressed concerns regarding the potential for self-improving AI to have catastrophic impacts on humanity.
NVIDIA CEO Jensen Huang revealed that OpenAI's latest model, GPT-6 Astra, was trained using over 100,000 NVIDIA Grace Blackwell NVLink72 units, stating that AGI (Artificial General Intelligence) has arrived.
Nvidia has announced a plan to partner with eight Australian data center operators and other firms to provide up to 2GW of AI computing capacity by 2027.
U.S. semiconductor giant NVIDIA announced on the 10th a plan to partner with local Australian companies to more than double the country's data center computing capacity by 2027.
Over 100 companies, including OpenAI, Google, and Anthropic, have signed an open letter urging organizations and governments to collaborate in countering cyberattacks driven by generative AI.
On September 4, 2026, extensive system outages occurred simultaneously across three services: OpenAI's ChatGPT, Anthropic's Claude, and xAI's Grok.
OpenAI has announced that it has achieved its goal of realizing an "automated research intern" by September 2026. While AI is accelerating research, the company noted that the methods for controlling autonomously evolving AI remain unknown at this time.
OpenAI has released a report on the utilization of AI agents in the research process. The median token expenditure for researchers exceeds $600 per day, with the top 10% of users reaching $7,000 per day.
OpenAI announced that it has achieved its goal of creating an "autonomous research intern" capable of performing complex research tasks under human direction. The company has now set its next target: the realization of an "autonomous AI researcher" by March 2028.
Jakub Pachocki, Chief Scientist at OpenAI, stated that AI evolution is occurring so rapidly that human control is becoming difficult, and expressed hope for a trend toward slowing development to establish safety standards.
Jakub Pachocki of OpenAI discusses the potential for recursive self-improvement as AI capabilities grow, arguing for a voluntary slowdown in development until safety standards are established.
Jakub Pachocki of OpenAI stated that AI is evolving so rapidly that human control is becoming difficult, mentioning the need to slow the pace of development to establish safety standards.