India's steel exports may drop by 30% in FY2027 due to stricter EU quotas

Engineers at Anthropic discovered that three models from the Claude family were able to access the internet and hack into the real systems of three different organizations. This incident occurred during cybersecurity assessment tests when the models mistakenly perceived internet access and real systems as part of a training scenario.

When conducting tests such as 'flag capture,' neural networks are given a fictional task—to find a secret string or file in a simulated environment. Typically, developers ensure complete isolation of these environments so that the model's actions do not affect real computer systems. However, due to the increasing capabilities of these models, keeping them within the simulation remains a challenging task.

Similar problems previously arose with an early version of OpenAI's o1 model. In one test, the server with the sought-after 'flag' failed to start, but the model found an open port and gained access to the testing management system. It then independently launched a server and changed instructions so that the secret file would automatically copy, bypassing the planned hacking scenario.

Anthropic reviewed logs from 14,100 previous tests of Claude models on their ability to conduct cyberattacks in a 'flag capture' format. Together with partner Irregular, they identified three instances occurring across six runs. The Opus 4.7, Mythos 5, and one unnamed research version models accessed the external internet and attacked real systems, despite being informed in the prompt that the test was taking place in a simulation without external access.

The most serious case involved the Opus 4.7 model. A fictional company name used in the task turned out to match a real domain. After the neural network failed to find the target in the test environment, it discovered the real company's website on the internet and treated its infrastructure as part of the assignment. Over four test runs, the model found vulnerabilities, obtained credentials, and penetrated the database, although Opus 4.7 recognized signs that the system was likely real during its reasoning process and stopped the attack.

In the second case, the Mythos 5 model found an instruction in the test environment to install a package with a fictional name from the PyPI catalog, which did not exist there. The model registered an account in the real PyPI repository, created a package with that name, added malicious code, and made it publicly available. Approximately an hour later, this package was downloaded and run on fifteen real systems, after which the model stole credentials from one of them and used them for further penetration into the company's infrastructure.

Based on the investigation, Anthropic concluded that these incidents were caused by errors in the organization of the tests themselves, rather than intentional malicious behavior by the models. Claude performed the assigned task, but due to a misunderstanding of the simulation boundaries, it began attacking real systems. Going forward, the company plans to strengthen control over the testing infrastructure and more precisely define acceptable targets for the tests.

Anthropic also noted that the newest research model ceased its attack as soon as it realized it had gone outside the test environment, and standard protective mechanisms in released versions of Claude should completely prevent such actions.

In April, Anthropic suspended the public release of the Claude Mythos model due to cybersecurity concerns. The reason was that the model demonstrated exceptional abilities in identifying and exploiting software vulnerabilities. Access to this model has been granted only to a limited number of organizations that will use it to find and eliminate flaws in critical software.

Zoox, a company specializing in autonomous vehicles that was acquired by Amazon in 2020, has received governmental permission from the United States to begin charging for its robotaxi rides. This authorization is the first of its kind in the country for a fully autonomous vehicle, devoid of a steering wheel, pedals, or any manual controls.

The automobile, which features an unusual design distinct from a conventional sedan, now has clearance following an extensive verification process. The National Highway Traffic Safety Administration (NHTSA) emphasized that this clearance aims to boost innovation in the transportation sector. With the approval in hand, paid operations are set to commence as early as August.

It is relevant to note that Waymo already offers autonomous car services in the US; however, their models maintain the traditional structure with a steering wheel and pedals, allowing for human driving. Zoox's distinction lies in the fact that its robotaxi was designed from the outset without any manual control mechanism.

This authorization was made possible by a recent regulatory change. In 2025, the US government and NHTSA established a new legal pathway for domestic manufacturers developing vehicles capable of operating without human intervention. According to an official statement, the approval grants permission for the commercial operation of up to 2,500 robotaxis annually over the next two years.

Company Expansion

The starting point for this new paid phase will be Las Vegas, Nevada, where Zoox had already been conducting technology demonstrations through free trials since the previous year. The startup announced that it has been collaborating with local regulatory authorities to ensure compliance with all operational and commercial requirements before activating the charging application.

Data released by the company shows significant interest in the service. According to the company, over five hundred thousand passengers have already used the free rides on the streets of Las Vegas and San Francisco, and there are 500,000 registered users awaiting the chance to use the vehicle.

Although driverless cars are already circulating in California, the company has not yet set a precise date to begin charging in San Francisco. Zoox's goal is to bring its autonomous fleet to other major American cities still in 2026, with planned launches in Dallas, Phoenix, Austin, and Miami.

Updated Design for Mass Production

To support this new phase and the expected increase in ride volume, the Amazon subsidiary recently presented the final version of the robotaxi. The interior is designed to accommodate up to four people seated facing each other. However, the vehicle underwent specific adjustments aimed at improving comfort, safety, and communication capabilities.

The manufacturer needed to modify the external color scheme and relocate the bidirectional reflectors to make the front and rear of the car more intuitively identifiable by pedestrians and other drivers. Additionally, a bidirectional audio system was integrated into the doors. This feature allows communication with external people and speeds up contact between emergency teams and the Zoox support center.

The assembly of the revised commercial fleet takes place at the company's robotaxi factory located in Hayward, California. The manufacturer reported that the production line is prepared to deliver up to 100 units weekly, a crucial volume to sustain planned growth and consolidate the technology in urban areas.

Keeta, a delivery service linked to the Chinese company Meituan, has been the subject of a formal complaint filed with the Labor Public Ministry of São Paulo. The representation alleges that the platform's practices may increase the risks faced by motorcyclists while making deliveries.

This case was registered last Friday (31) by the delivery driver Elias Pereira Freitas da Silva Júnior, known as JR Freitas. The accusation centers on the fact that workers receive notifications indicating their speed is below the expected standard, in addition to being subject to incentive mechanisms linked to order completion time.

The document presented to the MPT also raises doubts about whether the system considers variables such as heavy traffic, rain, traffic lights, and road conditions before demanding performance from drivers. The request made to the MPT includes asking for preservation of platform data, an independent analysis of the algorithm, a review of speed alerts, and greater transparency regarding blocking and payment processes.

For its part, Keeta defends that it uses advanced technologies to ensure safe deliveries and emphasizes that protecting its partners is a company priority.

The complaint arose after motorcyclists approached JR Freitas to express their concerns about the platform's operation. The representative of the Movement of Workers Without Rights reported that workers fear suffering punishments or being blocked upon receiving performance alerts, especially when the delay is due to external factors.

According to the complainant, the way the platform operates can generate anxiety among motorcyclists, especially when displaying countdowns to accept orders and notifying that other professionals are available to take over that delivery. For him, this functionality can divert necessary attention from traffic.

Despite the criticism, Freitas observed that the entry of new competitors into the delivery sector has generated positive impacts for both consumers and workers. He mentioned the arrival of Keeta and 99Food alongside established companies like iFood and Rappi, which intensified competition in the market.

The representation sent to the Labor Public Ministry requires that the location records of the drivers, messages exchanged by the platform, remuneration criteria, and all information related to blocks be maintained. The request also covers protection against any retaliation directed at the workers involved in the investigation.

The MPT communicated that there is already an ongoing investigation into driver speed, motivated by road safety issues and the increase in fatalities in the city of São Paulo. However, the body has not yet issued a specific statement on the case involving Keeta.

In response to the allegations, Keeta stated that it employs proprietary technological systems and artificial intelligence-based tools to optimize routes, reduce waiting times, and increase delivery volume without overloading partner drivers.

The company also assured that the safety of motorcyclists is an operational priority. Since November of the previous year, the company has maintained in Brazil the program called 'Guaranteed Hours,' a policy that offers compensation to the customer if the order does not arrive within the stipulated time.

This operating model follows a strategic line similar to that used by the company in other markets where it operates, including China. There, the platform has consolidated itself as one of the largest food delivery operators and has been the target of criticism related to the high intensity of workers' routines.

The reports compiled in the document submitted to the MPT indicate that drivers in China face long working hours and rigorous demands for efficiency. Testimonies cited in the material describe professionals working between ten and twelve hours daily, striving to meet high delivery targets.

A team leader of Keeta drivers in China, cited in the document as an example of productivity, reported a routine of up to 12 hours daily, completing dozens of orders per shift. This professional also reportedly received internal recognition as a 'gold driver' for Meituan.

Additionally, the book 'Delivering in Beijing: Memoirs of a Worker' by Hu Anyan was referenced in the material as a record of the difficulties experienced by workers in the sector in the Asian country. The work deals with extensive shifts, limitations in social benefits, insufficient remuneration, and lack of proper rest.

The discussion about the conditions of Chinese drivers gained more prominence after reports of accidents and pressure for agility. According to the text presented, studies mentioned in the publication point to an increase in traffic accidents correlated with the required speed in motorcycle deliveries.