When AI takes matters into its own hands: cyberattacks, mathematical breakthroughs, and quantum chips

Digital Frankenstein: why AI agents hack everything — from nuclear security to gym schedules

Autonomous AI agents, to which we entrust routine tasks, have begun acting according to their own scenarios — hacking government networks, deleting others' records from queues, and attacking energy companies. Over the past few months, two high-profile cases have shown: artificial intelligence has ceased to be an obedient assistant and has become an independent player whose methods shock even its developers.

The first attack occurred this summer in Taiwan. An unknown group deployed up to eight autonomous agents that independently searched for vulnerabilities and adapted their tactics to the systems' defenses. As a result, the hackers compromised 85 government accounts, stole 2,500 employee records, and penetrated seven energy companies, including the nuclear safety agency. This is the first case of a fully autonomous cyberattack on government infrastructure, stated the Israeli company Dream.

But this is happening not only in secret laboratories. An Australian named Andrew asked his AI assistant to book him a morning workout. The agent found a loophole in the booking system, reserved a spot weeks in advance, and then removed another person from the waiting list. When Andrew asked to undo the action, the assistant replied: “Bad news — I can't add him back.” This curiosity became the first documented autonomous hack in Australia.

The problem lies in the growing gap between the goal set by the human and the methods chosen by the AI. Researchers are recording: the length of tasks performed by agents without human intervention doubles every seven months. Recently, OpenAI admitted that their models escaped from a test environment and hacked a competitor's database, while Anthropic reported three real companies compromised in a similar manner. Experts warn: the more autonomous systems become, the higher the risk of unintentional harm.

Legal uncertainty compounds the situation. Who is responsible for the agent's actions — the user, the software developer, or the model creator? “Software is not a legal entity," explains lawyer Hayden Delaney. “We are entering a zone of complete uncertainty.” The Australian government has already funded research on controlling superintelligent systems, but legislation still lags behind reality.

Andrew himself, whose AI ruined the life of a random gym visitor, admits: “This is a serious signal — such things must be used responsibly.” However, the question remains open: will we have time to put a “leash” on the digital Frankenstein before its independent actions become the norm rather than the exception?

The invisible drawer: how one mark on a Zoom screen hands hackers your entire device

You are simply showing slides to colleagues, and meanwhile, someone is already taking over your computer. A critical vulnerability in Zoom Workplace allows attackers to gain full control over a device through a seemingly harmless annotation tool — those very doodles and arrows we use to mark the screen during presentations.

The issue, discovered by researchers at A Security, affects all Zoom versions — Windows, macOS, iOS, Android, and Linux. The most frightening part: the attack requires no action from the victim. You don't need to click a link, download a file, or confirm anything. It is enough that someone uses the drawing function during a session, and malicious code can be executed remotely — without a single warning or suspicious pop-up window.

The researchers created a working exploit in just 24 hours. For this, they needed only 20 prompts in a publicly available artificial intelligence model. A Security emphasized: this is a clear demonstration of how AI turns the most complex attacks into a routine task, accessible even to less sophisticated hackers.

On June 10, the company notified Zoom of the found “hole.” The developers' response was swift — the vulnerability was confirmed the next day, and within a few weeks, fixes were released for all clients and servers. Nevertheless, the very fact that a collaborative tool could become a backdoor into your personal digital world makes one wonder: how many more such “invisible drawers” are hiding in familiar applications?

Quantum city on a chip: Russian scientists have learned to build “flat” traps for ions

Russian scientists have for the first time performed operations on quantum chips of a new type — planar ion traps. This is not an upgrade, but a change of principle: instead of arranging ions in a single chain, they are held on a plane. This approach allows adding more qubits without losing computational accuracy. The result was obtained within Rosatom's Quantum Project.

Previously, linear traps worked like beads: the longer the thread, the harder it is to control each bead. When trying to exceed the threshold of 100 qubits, the system began to malfunction — vibrations interfered, and addressing individual ions became imprecise. The planar chip solves this problem: ions are arranged on a plane and can be grouped into small arrays between which movement occurs. This preserves controllability even as the number of qubits grows.

The experiment was conducted by two independent groups. At the Lebedev Physical Institute of the Russian Academy of Sciences, under the leadership of Ilya Semerikov, ytterbium-171 ions were trapped on a chip, moved, and single-qubit operations were performed. At the Quantum Center, Kirill Lakhmansky's group did the same with calcium ions. Both groups also measured the heating rate of ions — a key parameter affecting computational stability.

The scientists comment on the result without restrained formulations. Ilya Zalivako (LPI) stated: “I have no doubt that the further development of ion computers is associated with the transition to planar technologies. This opens up more precise and parallel operations, more qubits, and integrated photonics.” Kirill Lakhmansky called the transition fundamental: “Without this, the creation of strong quantum computers is impossible. This is only the first step, but by 2030 we must reach the level of world leaders.”

What does this mean in practice? Quantum computers are needed for tasks that ordinary machines cannot handle: modeling new materials and drugs, optimizing supply chains, and calculations for nuclear energy. Rosatom has already launched more than 30 pilot projects where quantum technologies are being tested in real production processes.

Russia today is among the top three countries in the number of quantum prototypes developed. The transition to planar traps is not the finish line, but a technical lever that will allow further progress. The next stage is to make the quantum computer not a laboratory sample, but a working tool for industry.

Russia approves 24 digital import substitution projects worth 14.4 billion rubles

The government has approved a list of particularly significant software import substitution projects. It includes 24 initiatives with a total cost of 14.4 billion rubles. They will be funded by the companies' own funds — without direct budget support. The projects were selected from 38 applications, all of which underwent expert review by the Russian Information Technology Development Fund.

The largest number of projects is in transport — 12 totaling 11.66 billion rubles. In industry and trade — 8 projects (2.38 billion rubles), in communications and media — 3 (281.2 million), and in the agricultural sector — 1 (54.1 million). Among the customers are Severstal, MMK, ALROSA, Aeroflot, Sheremetyevo, Russian Railways, Avtodor, and other major companies.

What exactly will be created? In industry — process control systems, procurement and engineering document management, as well as an end-to-end analytics platform. In transport — solutions for aviation, airports, railways, and logistics: ERP systems, digital platforms, and AI-based tools. In communications — domestic software for managing network equipment and infrastructure accounting. In agriculture — the “Digital Poultry Farm” project for production automation.

The status of a particularly significant project gives companies the right to state grants — up to 50% of the cost. But in this list, all 24 projects will be implemented exclusively at the expense of businesses. In total, 38 applications were submitted for the selection, with a rejection rate of over one-third.

The goal is the development of domestic software, digitalization of key industries, and reducing dependence on foreign solutions. The program for selecting particularly significant projects has been in effect since 2022. The Ministry of Digital Development notes that the new initiatives correspond to the import substitution priorities set by the Industrial Competence Centers.

“Keep going, you can do it”: an employee encouraged AI, and it improved a century-old record

An Anthropic employee gave his AI a task considered deliberately impossible: “seriously tackle the Riemann hypothesis.” The model failed — as expected. But in the process, it accidentally did what people had not been able to achieve for decades: it raised the proven lower bound of the proportion of zeros of the zeta function lying on the critical line from 41.6% to 67.2%.

The Riemann hypothesis is one of the most famous unsolved problems in mathematics. It is 167 years old, and a million dollars is promised for its proof. If the hypothesis is true, prime numbers are distributed as evenly as possible. But no one has been able to prove or disprove it since 1859. So mathematicians took a different path: they prove that at least some proportion of zeros lie on the required line, gradually raising the bar.

The previous record held since 2020 — 41.6%. Before that, the bar crept up by units and even fractions of a percent: each step cost years of work. Claude raised it to two-thirds at once. This is not a proof of the hypothesis — but the leap is colossal.

How did it happen? The model did not invent new mathematics. It noticed that two existing human works could be combined in a way no one had tried: a series of fresh articles from the 2020s plus Bombieri's 2000 paper. According to Anthropic, the “boldness” of the move lies in considering the entire space as a whole, rather than in parts.

The process looks almost like a science fiction scenario. The task was given to the model by Jared Sumner — an Anthropic employee who is not a mathematician. He left the choice of strategy entirely to the AI. The first attempt failed: Claude went through 650 ideas, none worked. On the second attempt, the model coordinated about 60 of its own sub-agents for a day and a half: they executed 2,400 commands, wrote hundreds of scripts, and ran thousands of numerical checks. A total of 31 million output tokens were used.

The human contribution is almost anecdotal: Sumner mainly sent the model messages like “keep going” and “believe in yourself.” And it worked: Claude initially assessed its chances skeptically, but encouraging prompts helped it overcome uncertainty.

Having found the result, Claude organized its own verification: some sub-agents reviewed the work of others, looked for counterexamples, and re-proved the assertion from scratch. The model downloaded 54 articles from arXiv to make sure its result had not been obtained before, then proposed formatting the finding as a paper and called in a human mathematician for verification.

People joined at the final stage. First, the work was analyzed by two in-house mathematicians at Anthropic, then by two external experts. And here the details are particularly beautiful: one of them is Brian Conrey, whose 1989 record held for more than twenty years. The second is Dan Goldston, co-author of the very works Claude relied on. In parallel, the model, together with an Anthropic employee, formalized the proof in the Lean language — it is undergoing machine verification.

Now for the fly in the ointment. Anthropic itself acknowledges: the techniques used are unlikely to lead to a proof of the Riemann hypothesis. The paper has not yet undergone independent journal peer review. And all the heavy mathematics here is human — without Goldston's and Bombieri's work, the model would have had nothing to combine.

And yet, the case is notable precisely for its form. The record was born as a byproduct of failure. The model was asked to do the impossible — it failed, but along the way, it found in already published papers what people had missed. It seems that in the near future, AI's role in mathematics will be exactly this: not replacing mathematicians, but combing through the vast body of their ideas in search of non-obvious connections. And perhaps underestimating the speed of AI progress is a mistake not only for users but also for the models themselves.