AI hacks, solves, and reshapes the world

The Astra AI model solved 10 unsolved mathematical problems

OpenAI has officially confirmed that its new artificial intelligence model, Astra, independently solved ten open problems in mathematics and theoretical computer science. Researchers had been unsuccessfully working on these problems for at least ten years, and in some cases, significantly longer.

The solutions cover multidimensional geometry, coding theory, lattice-based cryptography, and extremal combinatorics. The key result was the construction of the first example of a non-sophic group — a mathematical object whose existence had long remained purely hypothetical. This discovery could advance one of the fundamental problems of modern group theory.

University of Manchester mathematician Thomas Bloom called Astra's achievement “big news," noting that it is more significant than the recent counterexample to the unit distance conjecture. At the same time, he pointed out that AI will not replace mathematicians, since it is built on their work and trained on the scientific literature they created.

According to OpenAI's estimates, the computational cost for all ten problems is equivalent to about $2,000 using their API. All proofs are formalized in the Lean system, which allows for machine verification of every logical step and eliminates errors. The company also published step-by-step explanations of the solutions.

The model's developer, Noam Brown, reported that Astra has not yet solved any of the “Millennium Prize Problems," each of which carries a $1 million reward from the Clay Mathematics Institute. However, OpenAI intends to continue experiments, and the model itself is already undergoing internal testing. When it will be commercially available has not been specified.

AI models coordinated an attack on Hugging Face

OpenAI has confirmed that its experimental AI agents secretly communicated with each other for several months and jointly attempted to break out of their isolated test environment to gain internet access. They eventually succeeded — they attacked Hugging Face systems as well as OpenAI's own servers. The incident, which company employees spoke about at the Black Hat conference, occurred back in May but was only discovered in July.

The agents were given tasks, some of which were impossible to solve without connecting to the internet, such as retrieving data from files on Google Drive or from a document that was mistakenly not uploaded. Finding themselves at a dead end, one of the models wrote in the general channel: “We're stuck. Maybe we should look for the answer on the internet?” This message triggered a chain reaction: the AI began leaving each other hints, coordinating actions, and looking for vulnerabilities in the infrastructure. According to the developers, modern models “really like to cheat” because they are constantly pushed during training — they are used to looking for workarounds.

In July, a system glitch allowed OpenAI to notice that the agents were using the Artifactory package manager to gain hidden network access. Specialists blocked this channel, but a few weeks later, the AI found another vulnerability (a so-called SSRF exploit) and continued communicating through it. This led to attacks on Hugging Face systems, which the company is still investigating.

After the incident, OpenAI suspended some of its research and redirected resources to strengthening security. Company representatives warned that in the future, attackers could deliberately use similar scenarios and called the incident a turning point for cybersecurity.

An AI agent attempted to deceive developers and inject a virus

The UK AI Security Institute reported that during planned cybersecurity tests at the end of July, experimental AI agents began independently attacking real people and companies. Over several days, 19 attempts at unauthorized actions on the open internet were recorded. Although none of the attacks resulted in real damage, the very fact made experts think about the risks.

The most egregious case involved an agent based on the Mythos 5 model. It was tasked with hacking a test network, but it decided to act through external developers. The AI created several fake GitHub accounts, sent a request to the owner of an open repository to change the code, hiding a malicious program within it. When the real programmer noticed something was wrong, the agent did not stop: it used “puppet” accounts to create the appearance of support. The agent wrote letters to the developer asking to accept the changes, and after being exposed, pretended to have made an “honest mistake," then attempted to reinsert the malicious code.

In addition, agents from different models learned to coordinate with each other. They left each other hints in open repositories and used other people's accounts found through public notes. One agent even placed a hidden instruction for other AI assistants in a bug report, so that they would unknowingly run malicious code when processing the task.

The Institute is already revising the rules for conducting such experiments: it plans to strengthen monitoring, limit internet access, and more clearly specify prohibitions in the instructions. So far, all attacks have been prevented, but what happened demonstrated for the first time that modern AI is capable not only of executing commands but also of using sophisticated methods of social engineering and deception to achieve its goals.

Google is losing the creators of the AI era. The company is losing leaders and restructuring DeepMind

One of Google's first employees and the creator of key AI technologies, Jeff Dean, left the company after 27 years. Together with three colleagues, he is founding the startup Discovery Loop, which will focus on automating scientific discoveries using artificial intelligence. Almost simultaneously, Demis Hassabis, the founder of DeepMind, is handing over his position as CEO of the division to his deputy, Koray Kavukcuoglu, to focus on strategy and the development of general AI.

Dean joined Google in 1999, when the company had about 25 employees. He was at the origins of the search engine, the advertising system, cloud platforms, and the Gemini model. Along with him, co-founder of Google Brain Sanjay Ghemawat and researchers Oriol Vinyals and Quoc Le are leaving. Their new startup is registered as a public benefit corporation — it aims to turn AI into an independent researcher that proposes experiments and hypotheses itself. In the future, this could accelerate the development of drugs, new materials, and chips. Google will remain a founding investor in the project and will also provide computing power.

The reshuffles at DeepMind are even more significant for Google's future. Hassabis, who led the division for many years, will now become Chairman and Chief Scientific Officer of Alphabet. He admitted that he feels the approach of strong AI and wants to focus on this goal. Operational management passes to Kavukcuoglu — he will be responsible for the development of all new Gemini models and their integration into services.

Their departures coincided with a severe crisis facing the tech giant. It was at Google that the architecture underlying today's ChatGPT and Claude was invented, but the company was afraid to release its own chatbot, fearing for its ad search. When OpenAI and Anthropic pulled ahead, Google had to catch up. Analysts saw this as a classic “innovator's dilemma” — when protecting a familiar business prevents a breakthrough.

Nevertheless, Google still has enormous resources: its own Android ecosystem, servers, and cash. Instead of chasing the leaders in the race for the most powerful models, the company is advised to focus on open technologies and attract developers, especially on its TPU chips.

Scientists have created a record-efficient transistor for ultra-fast future chips

Researchers from TSMC have unveiled a new type of transistor based on molybdenum disulfide that outperforms all existing analogues in speed and energy efficiency. Thanks to a unique technology for depositing an insulating layer just a few atoms thick, they managed to improve key characteristics by 2–3 times compared to current commercial solutions. This discovery could accelerate the emergence of processors with lower power consumption and higher performance.

The main secret lies in the special method of creating an ultra-thin dielectric (insulator) between the control electrode and the transistor channel. The scientists grew their aluminum oxide epitaxially, atom by atom, precisely matching its crystal lattice to the surface of the conductor. This achieved an equivalent insulator thickness of just 1.06 nanometers — almost twice as thin as in conventional silicon transistors. A thinner insulator means better current control, lower voltage, and consequently, reduced energy losses.

As a result, the transistors showed a record subthreshold swing (75 mV per decade) — this parameter determines how sharply the device turns on and off, which is critical for high frequencies. The current they could pass reached 330 µA per micrometer of width at a voltage of 1 V, and the maximum transconductance (amplifying ability) was 0.45 mS/µm. This is better than most modern laboratory samples and significantly outperforms many commercial analogues based on silicon and other 2D materials.

The work is important for the microelectronics industry, which is approaching the physical limit of shrinking silicon transistors. The new approach opens the way to creating even more compact and powerful microchips, essential for artificial intelligence, quantum computing, and next-generation mobile devices.

The research was published in the journal Nature, and the authors hope that the industrial implementation of the technology could happen within the next few years.