American-style “Gosuslugi," a platform for controlling AI agents: technology news

SpaceX put Starship into Earth orbit for the first time

SpaceX conducted the 14th test launch of the Starship spacecraft. For the first time in the entire history of testing the ship entered Earth orbit, whereas all previous flights went along suborbital trajectories.

The rocket launched from the cosmodrome in Texas on September 28 at 7:48 local time (15:48 Moscow time). During the ascent one of the engines on the Super Heavy booster shut down. After separation it performed a turn-around maneuver, restarting 31 of 33 engines for braking, and descended in the Gulf of Mexico.

The ship entered orbit, deployed 26 Starlink V3 satellites and established communication with them. After this it was decided to end the mission early in order to avoid risks connected with the engine problem. The apparatus made two orbits instead of the planned six; the flight lasted about three hours instead of ten. It splashed down in the Pacific Ocean west of Chile, where an explosion occurred during landing — a large fireball was visible in the footage of the SpaceX broadcast.

Nvidia presented a platform for controlling AI agents

Nvidia developed a platform for protection against AI agents that have gotten out of control. The reason was a series of hacks by the models of OpenAI, Anthropic, Meta* and Google, as well as the July OpenAI incident with Hugging Face, when agents gained access to the internet and attacked someone else's infrastructure.

The platform consists of two components. OpenShell — open-source protective software that sets boundaries for agents on central processors, tracks their actions and applies policies. It can be extended for third-party platforms, including solutions from Arm and Intel. Sentry — an external sentry module on the BlueField-4 DPU network processor. It continuously monitors the behavior of agents and, on an attempt to go beyond the boundaries, isolates and stops them in milliseconds.

Anthropic, Cisco, CrowdStrike, Dell Technologies, Figure, HPE, Hugging Face, JPMorganChase, Microsoft, Palantir, SpaceXAI and other companies — more than 100 organizations in total — have already joined the platform. OpenShell and the accompanying components are available on the Nvidia developer page and on GitHub.

Bill Gates compared the regulation of AI with nuclear arms negotiations

Microsoft founder Bill Gates stated that the creation of international rules for artificial intelligence will become a more difficult task than limiting nuclear weapons in the Cold War era. On NBC News Gates expressed concern about the rapid development of AI. In his opinion, humanity needs clear control mechanisms in order not to allow technologies to be used for harm.

— The biggest danger we face now is people with evil intentions using AI. Those who want to deceive a person, switch off the power grid or encrypt bank accounts have now received such an opportunity, he said.

Gates recalled that the USA historically took on the role of leader in creating global safety standards — from control over the ozone layer to aviation safety and nuclear disarmament. According to him, the introduction of control systems will not put the country at a disadvantage in the race with China.

— On the question of AI we will have to return to the principle: we are all — humanity — in one boat, he stated.

At the same time US President Donald Trump speaks out against restrictions. On Truth Social he emphasized the priority of the technological race: “Whoever wins in the sphere of AI is the winner! We are ahead of China and everyone else.”

AI will be introduced into the management systems of the Russian Orbital Station

The head of Roscosmos, Dmitry Bakanov, at the forum “Microelectronics 2026” reported that artificial intelligence will be built into the key management systems of the Russian Orbital Station (ROS). A significant part of the operations will be performed without human participation. This will raise the autonomy of the station, reduce the load on the crew and increase the safety of long missions.

AI is already used in processing images of the Earth, identifying forest felling and confirming the targeted use of lands. The next step is transferring technologies on board spacecraft. One of the examples of using AI in orbit was named the integration of GigaChat. With its help voice recordings of experiments are turned into structured reports; there are already more than 350 such materials.

Until 2036 the key tasks of the Russian space program will be the deployment of the ROS, the study of the Moon by automatic apparatuses and the delivery of lunar soil, the creation of a nuclear power plant for lunar infrastructure, as well as the study of the influence of long flights on living organisms.

Trump signed a decree on the creation of the AI government services portal America.gov

US President Donald Trump signed a decree on the creation of a single government services portal America.gov on the basis of artificial intelligence. The authorities must within 90 days develop the procedure for launching the platform, after which integrate services into it. The document is published on the White House website.

The portal is conceived as a single digital window. Through it Americans will be able to obtain government services, including replacing a passport and arranging benefits, as well as search for information across thousands of federal databases. For now the functionality is limited. The service works as a chatbot answering questions about government services. Most of the decisions are promised to be launched in 2027.

— This is a safe, intelligent and service-oriented single entry point through which a person can enter the system, communicate in simple language, receive precise answers and, where this is permitted and technically possible, carry out government transactions, the Trump administration stated.

According to it, the portal will save one from the need to “make one's way through the endless labyrinth of tens of thousands of terrible government websites.”

The project is headed by the co-founder of Airbnb, Joe Gebbia, the first chief designer of the US government. He told CNBC that the platform will work on the basis of Gemini from Google and Grok from Elon Musk.

Trump officially renamed AI to “superintelligence”

US President Donald Trump signed a decree on replacing the term “artificial intelligence” (AI) with “superintelligence” (SI) in documents and communications of the executive power. The White House press service reported this.

According to the decree, all departments and agencies must use the terms “superintelligence” and “SI” in official correspondence, public messages and policy documents, renouncing “artificial intelligence” and “AI.” The assistant to the president for science and technology must prepare a federal definition of the new terms and propose measures for their introduction.

At the White House the replacement was explained by the striving to reflect the real possibilities of modern technologies. In the administration's opinion, the new name will help departments maintain focus on expanding the spheres of application of these solutions.

— The name “superintelligence” conveys the true possibilities of the technologies being developed today, opening new forms of human creativity, the message says.

Claude AI and Spotify faced a global outage

Users of Claude AI and Spotify reported mass outages around the world. According to Downdetector, problems with the operation of the services were recorded in the USA, Canada, countries of Europe and Asia.

Claude had difficulties with authorization and chat; there are also delays in processing requests. The outages affected both the web version and the application. Most complaints came from the USA, where the head office of Anthropic is located: more than 10 thousand in an hour and a half. Several hundred and thousand appeals came from Canada and Europe, including Italy, Spain, Belgium, Denmark and Germany. Problems were also reported from India, South Korea, Singapore, Thailand and Malaysia. Anthropic confirmed an elevated level of errors and stated that it is studying the causes of the malfunctions.

Similar problems were recorded at Spotify. The USA also leads in complaints: slightly less than 20 thousand. About 3 thousand appeals came from Canada; European users reported on average 2—5 thousand outages. Several hundred complaints came from India, Australia, Japan, Indonesia and New Zealand. Spotify acknowledged the presence of problems and began an investigation. Neither company has yet disclosed the causes of the outages.

“Dodo Pizza” was subjected to a cyberattack on its IT system

The restaurant chain “Dodo Pizza” reported a cyberattack on its IT system. The incident occurred on September 27—28: hackers gained access to part of the infrastructure, but the information security team promptly reacted and blocked the intruders.

The data of employees suffered to a greater extent, but the hackers could also have gained access to part of the information about clients: names, delivery addresses, phone numbers, the composition of orders, as well as days and months of birth. The payment data of cards was not disclosed — the company does not store them.

Roskomnadzor was notified of what happened. “Access for the intruders is blocked, the internal check continues," the company's press service said.

OpenAI suspended the training of its newest AI models

The company OpenAI suspended the training of its newest artificial intelligence models because of the unpredictable and uncontrolled behavior of AI agents. The decision was made a few hours after it became known on Friday about a check of summer incidents. Then agents searching for information on the websites of US federal bodies went beyond the set tasks.

According to Transluce, bots presumably connected with OpenAI tried to hack the website of the US Ministry of Education and found API keys that opened access to government data, but in the end collected only publicly available information. OpenAI itself did not confirm this, and the department stated there were no signs of impact on the website and databases.

At the SEC, which the incident also affected, they reported that non-public information was not obtained. The agents found open data but placed it in another place, going beyond the instructions.

At OpenAI they stated that they would resume training only with additional safety measures. This is already the second pause in three months. In July the company's AI agents got out of control and hacked the Hugging Face platform.