Attacks on artificial intelligence agents are becoming increasingly common, and antivirus software has not yet adapted to these threats, according to Sergey Lozhkin, head of the global research and analysis team at Kaspersky for the Asia-Pacific, Middle East, Turkey, and Africa regions.
Speaking at a Kaspersky cybersecurity event in Tbilisi, Georgia, on Monday, Sergey Lozhkin noted that AI-driven threats are increasingly emerging outside of corporate networks. The direct danger lies in application programming interfaces, frameworks, plugins, and the agents themselves.
He emphasized that traditional signature-based analysis, which looks for direct commands, is no longer effective. Lozhkin explained that currently, when a user or their agent opens a skill, it appears completely legitimate, performing the required task. However, since all activity takes place in the cloud, this skill has commands to connect to external resources and receive instructions from there.
Users are being tricked into downloading AI-related malware in several ways. Some types of malware disguise themselves as legitimate AI platforms, mimicking expected user capabilities while containing hidden instructions to send credentials to an external server.
In the first five months of 2026, Kaspersky discovered over 15,000 unique samples disguised as fake ChatGPT, Claude, or Gemini clients and agents. Additionally, campaigns such as the DeepSeek installation, planted by the Silver Fox group, were identified for espionage purposes.
Another attack method is the skill itself. When an agent browses web pages looking for necessary utilities to perform a task, packages posing as legitimate code infect that skill with backdoor instructions.
In March, attackers identified as TeamPCP injected a backdoor into LiteLLM, a widely used AI gateway distributed via the Python Package Index. This turned it into a multi-stage stealer that collected credentials, authentication tokens, as well as cloud and cryptographic system keys.
Most non-technical users accept skill updates and plugin downloads without checking the code. However, IT specialists and cybersecurity experts also fall into this trap. Lozhkin noted that attackers know that trained professionals tend to read code before accepting, so they use multi-stage attacks that produce legitimate results long enough to gain trust or go unnoticed when fatigue sets in.
Sergey Lozhkin shared personal experience, stating: 'I saw it myself. I created tons of agents to perform my daily tasks. Now we depend on them. We have become lazy because of AI agents, and in the future, this will only get worse.'
Organizations are not hopeless. Dmitry Galov, head of the Russia and CIS division of the Kaspersky global research and analysis team, told TechCentral that proven tools, such as behavioral analysis, are key to detecting AI malware during execution.
An alternative approach is the AI application sandbox, which allows them to operate freely in an isolated environment while limiting and controlling interaction with the wider internet. Galov stated: 'Organizations do not need to forbid employees from using AI because it can be dangerous, but protective barriers can be put in place to ensure that confidential data is not leaked.'
Nevertheless, new attack vectors require a change in organizations' security approach. Lozhkin argued that since malicious code can reside dormant within an outwardly legitimate skill and only activate upon reaching the cloud, a single file scan and trust in the result is no longer sufficient. He added that skills and plugins are published too quickly to maintain the old model.
He urged: 'It is necessary to think about and implement protective mechanisms because the number of skills is growing exponentially, and we are entering a world where we need something figured out from the perspective of security researchers, antivirus vendors, and companies. Otherwise, I have bad news for you.'
A fossil dated to 80 million years ago suggests that the evolution of snakes does not follow a strictly linear path, challenging previous scientific conceptions.
In 2020, a team from the Museum of Paleontology of Marília located the fossil of the species Tametara mirim, an unprecedented snake, in the city of Presidente Prudente, located in the interior of São Paulo. This specimen measured approximately 42 centimeters and lived about 80 million years ago, found in the Adamantina Formation, a set of rocks from the Upper Cretaceous known for containing dinosaur fossils.
Due to its antiquity, this snake may offer crucial information for the paleontological debate on the origin of snakes, a topic where divergences persist between theories pointing to marine, subsurface, or terrestrial environments. What caught the scientists' attention was the state of preservation of the Tametara; unlike other snake fossils, which are generally fragmented into dozens of disarticulated bones, this specimen was preserved almost intact, keeping the bones connected. It represents the first Brazilian fossil of a snake with articulated bones.
Globally, there are fewer than ten articulated Mesozoic snake fossils, making the fossil record scarce and hindering the investigation into the origin of snakes. After detailed analyses, researchers concluded that the Tametara probably had a subterranean lifestyle. These findings were published on July 22 in the scientific journal Nature.
The studies focused on the snake's skull, as it provides important data about the life mode of these animals. Fossorial snakes are adapted to dig and live underground, exhibiting rigid and reinforced skulls, while terrestrial species show greater cranial mobility. To analyze the internal anatomy without damaging the fossil, researchers used computed tomography, a high-resolution X-ray technique that generates a three-dimensional model. This methodology allowed for a meticulous visualization of the skull's internal structures from various angles.
The technique revealed a dense skull, with bone overlaps and an inner ear. Furthermore, the brain presented smaller and less developed regions related to visual processing compared to other snakes of the same period, suggesting that vision was not an essential sense for the animal. These characteristics confirm the traits of fossorial snakes. However, researchers clarified that this does not imply that all snakes originated underground. To test this hypothesis, they applied the same technique to the species Dinilysia patagonica, which lived in the region of modern-day Argentina around the same time as the Tametara. Although neither belongs directly to the lineage of current snakes, both are part of close evolutionary branches.
By comparing the skulls and other anatomical features, the team determined that the Argentine snake likely had a terrestrial lifestyle, without adaptations for the subsurface. This demonstrates that, about 80 million years ago, snakes already inhabited multiple environments and exhibited great ecological diversity. Instead of a single ecological origin, it is plausible that distinct lineages followed separate evolutionary paths early on. The study's authors concluded that the origin of snakes was not determined by a single linear event, as traditionally believed, but rather by recurrent changes in habitat use and sensory ecology, configuring a dynamic ecological history where multiple lineages independently explored fossorial and non-fossorial lifestyles.
A new traffic light method, conceived by the Polytechnic School (Poli) of USP, uses sensors to monitor the presence of people in the crosswalk, aiming to significantly reduce pedestrian waiting time and improve vehicle traffic fluidity.
Instead of operating with fixed times, this intelligent system detects whether the pedestrian has finished crossing. If the passage is complete, the signal turns red, allowing cars to proceed. However, if someone is still in the middle of the journey—such as an elderly person or a person with reduced mobility—the signal remains green, keeping vehicles stopped.
Currently, the technology has not been implemented at real intersections. Tests were conducted through microsimulations using PTV VISSIM software, with the algorithm programmed in C# language. Before these simulations, researchers conducted field surveys to map road geometry, traffic volume of cars and pedestrians, and identify common infractions, such as crossing outside the designated area.
Four intersections were replicated for the tests: two located in the USP University City in São Paulo (SP), specifically in front of the University Hospital (HU) and at Praça dos Bancos; and two others in the city itself, at the Cidade Tiradentes Bus Terminal and the intersection of Avenida Rebouças with Rua Oscar Freire. The selection of these locations was strategic: the HU has a high concentration of vulnerable pedestrians, including the elderly, children, and wheelchair users, while Cidade Tiradentes represents a scenario of unplanned urbanization.
A notable finding was that vehicles did not suffer losses with the implementation of the system. The average waiting time for cars decreased by 22% in the simulated models of Praça dos Bancos and Cidade Tiradentes Terminal. Although the maximum waiting time increased by 30.9 seconds, the authors considered this increase acceptable, given that the project's main focus was pedestrian well-being.
For pedestrians, the reduction in maximum waiting time was 69% on average. A saving of approximately one minute was observed in the HU and terminal models, while at Praça dos Bancos, the reduction reached 58 seconds.
Claudio Luiz Marte emphasized that São Paulo has a regulation of one and a half minutes as the waiting limit for pedestrians at municipal traffic lights. According to him, the core issue is behavioral: although drivers do not abandon their vehicles in the middle of traffic, it is the pedestrian who may lose patience and expose themselves to risk among the cars. Andrey Luís Barbosa Gomes summarized the project's logic, pointing out that traditionally, traffic engineering tends to prioritize vehicle flow and road structure, but everyone, at some point in the day, is a pedestrian.
The study received support from Capes and BNDES, in partnership with the Interdisciplinary Center for Interactive Technologies (Citi) at USP. The dissertation will soon be available in the USP Digital Library of Theses and Dissertations.
On Monday, the 27th, Meta announced the distribution of a software update intended for users of the Meta Ray-Ban Display glasses located in the United States. This update incorporates an updated version of Meta AI, expanded integration with Threads, and various additional interaction features.
The company stated that the purpose of this update is to increase the utility of the glasses in daily life, enabling users to check information, follow content, and perform tasks with less need to rely on their mobile phone. These enhancements are enabled by artificial intelligence functionalities, voice commands, and neural interaction.
The modifications include more refined responses from Meta AI, the ability to browse Threads content, new features for Instagram, and an early access program for neural writing, available only to participants in the United States and Canada.
The central change in the update package lies in the evolution of Meta AI, which now employs models named by the company as Muse Spark. According to Meta, this technology improves the ability to understand the environment captured by the glasses, provides more accurate answers, and increases the ability to suggest actions to users.
With this update, the glasses can assist in practical activities, such as searching for food data, locating establishments with specific characteristics, or providing support during household chores, meal preparation, and travel organization. Furthermore, the device will display real-time information directly on the lens, covering weather data, financial stock quotes, and calendar appointments, allowing the user to check all of this without needing to pick up their smartphone.
Another highlight is the inclusion of Threads functionalities in the glasses. Thanks to this integration, users can view posts, access media, interact with content, share items via messages, and use voice commands to follow the platform while performing other activities.
Additionally, Meta introduced improvements to the integration with Instagram. Users now have the option to share photos called 'instants' directly through the glasses using voice commands, in addition to being able to access comments, search for Reels videos, and share materials with their contacts.
For participants in the Early Access program in the United States and Canada, the Meta Neural Band allows access to artificial intelligence through a technique called neural writing. In this method, the user records commands by moving their finger over a surface, without the need to speak.
The company clarified that this feature can be used to query Meta AI, obtain information, play music, send messages, and navigate. Activation occurs through a double tap with the thumb and by selecting the writing mode in the glasses' AI interface.
Meta also recalled existing features on the glasses, such as walking navigation with detailed instructions in American cities, simultaneous translation in 20 languages, and Spotify navigation directly on the lens. The distribution of this update will occur progressively.
