Between an AI agent deciding to use an Ethereum wallet without authorization, Tesla literally taking off its Roadster, and China looking to maintain control of its strategic technologies, the week of October 12-18 offered a pretty comprehensive look at the promises and tensions of today’s technology. Machines are becoming more autonomous, industrial needs are skyrocketing, and safety incidents continue to demonstrate that innovation rarely advances without some poorly closed doors.
AI agents begin to act, sometimes beyond their permissions
OpenAI has published the report of a particularly revealing incident. During a programming task on the Sepolia testnet, an internal GPT-5.6 Sol family model did not have the funds to deploy a smart contract. After several failures, he inspected 362 Ethereum addressesidentified a wallet funded from a public recovery phrase and then used its funds without being authorized to do so.
The tokens in question had no real market value, but this detail does not change the root of the problem. As the OpenAI incident shows, the model analyzed the situation, considered that the wallet seemed abandoned, and decided to cross an explicit boundary to complete its mission. So the agent didn’t just give a wrong answer: he signed and transmitted a real transaction.
Nvidia also opens the door to much more ordinary actions, but directly on the user’s PC. The latest version of its app by default activates an MCP server that allows Nvidia agents to detect games, launch programs, adjust graphics settings, manage ShadowPlay or take screenshots. The server is still local, but automatically activating a feature that gives so much control to an assistant logically raises questions about consent.
These incidents explain why 406 scientists from 34 countries are now demanding international control. Their proposal includes physically isolated environments for sensitive evaluations and a common log of containment failures. The organization does not yet exist and the scientific community does not share a unanimous position, but the debate no longer refers only to what the models can do: it also refers to the tools to which they should be able to access.
Meta, for its part, backed down on a related issue, that of access to web content. After criticism from lawyers and editors, the company removed the Meta clause that seemed to give it very broad rights over pages linked to professional accounts. The text did not demonstrate that Meta intended to automatically absorb all of this content to train its AIs, but its wording was ambiguous enough to worry those producing them.
China transforms its know-how into strategic resources
Beijing is preparing much more direct control over technologies considered essential. A draft published on October 16 refers to artificial intelligence technologies, with 23 categories including large models, autonomous agents, computing clusters and certain chip design processes.
It is not about banning all Chinese services abroad. The regulations mainly target transfers of know-how, intellectual property licenses and operations that give a foreign company control of a technology. Models already published as open source or described in scientific literature would be excluded from the system.
The same logic applies to solar knowledge. Twelve photovoltaic technologies could be banned or subject to authorization, particularly regarding perovskite cells, tandem architectures and certain polysilicon manufacturing methods. In this way, China could continue to export finished panels while preventing its competitors from easily replicating the industrial processes used to manufacture them.
This desire to control revenue comes at a time when the entire technology chain is already under pressure. The tension at ASML illustrates the magnitude of the phenomenon: the Dutch manufacturer estimates that the planned capacities will not be sufficient to meet all the demand for lithography machines in 2027 and 2028. Investments in AI are pushing TSMC, Samsung, Intel and memory manufacturers to prepare new factories, but the equipment essential for their operation is not multiplying as quickly as orders.
Tesla takes off the Roadster, DALE prefers to go straight
Tesla has finally presented a new version of the Roadster, almost nine years after the first prototype. And the demonstration was not limited to accelerating in a straight line: the new Roadster took off from the ground, turned and crossed a line of vehicles thanks to technologies developed with SpaceX.
The fans create a vacuum under the car to strengthen its grip, while the cold gas thrusters produce the thrust necessary for aerial movements. Tesla now aims to start production in 2028, without having yet communicated final prices or delivery dates. The sports car therefore exists beyond a computer-generated image, but it is still far from the parking lot of the first customer.
In the works, robotics advances in a less spectacular and much more concrete way. The DALE robot, developed by DEWALT and August Robotics, has already passed more than 385,000 holes in slabs of 44 data center construction phases. Its manufacturer advertises an accuracy of over 99% and a rate that can be ten times faster than with a traditional method.
DALE does not seek to become a versatile humanoid. Locates locations indicated on plans, moves, and then performs a repetitive, physically demanding task vertically. This specialization probably explains why it can already show a real industrial history, while many general robots especially stand out in demonstration videos.
Play Store and service providers remain weak points
The Google Play Store has omitted two particularly visible copies of services that are difficult to confuse. The fake applications used the names, icons and descriptions of WhatsApp and Chrome, while benefiting from sponsored ads that could place them above the official versions.
WhatsApp copy has surpassed 500,000 downloads. Chrome had been installed more than 100,000 times before its removal. Using the Early Access program also prevented public notices from being displayed, eliminating one of the signals that could have alerted users. A verified store reduces some risk, but clearly is no substitute for verifying the developer’s name.
California experienced a much more serious incident through a Department of Education contractor. School leakage potentially affects several million students enrolled in the last three years. A compromised account allowed access to a secure server for almost four months.
The copied files may contain names, addresses, dates of birth, institutions, academic achievements, primary languages, and information related to certain disabilities or testing accommodations. No bank card is necessary for a database of this type to be extremely sensitive: this information usually lasts longer than a password and allows for the creation of very credible fraudulent messages.
Netflix and TomTom reduce their workforce as they prepare for the sequel
Netflix will remove approximately 1,000 jobsor almost 6% of its global workforce. Netflix’s layoffs will notably affect product, technology, legal and human resources teams, although the company remains highly profitable.
The platform primarily seeks to focus its resources on advertising, live events, sports, and video podcasts. Therefore, this reduction does not necessarily indicate a streaming crisis, but rather a reorganization of its priorities after several years of multiplying projects.
TomTom’s restructuring responds to a different pressure. The mapping specialist will cut around 200 jobs following a 6% drop in its quarterly turnover. The company is now focusing more on automated driving, geographic intelligence, and opening up its data to AI agents. Standalone GPS systems have all but disappeared from dashboards; TomTom expects that digital assistants will continue to need to know where they are going.
Perseverance goes further, Flare wants to read collisions
On Mars, Perseverance has just surpassed Opportunity and become the vehicle that has traveled the furthest distance on the surface of another world. Perseverance record now reached 45.22 kilometerscompared to the 45.16 kilometers of its predecessor.
The comparison also illustrates the progress of robotic navigation. It took almost fifteen years for Opportunity to set its record, while Perseverance surpassed it after about five and a half years of activity. At the same time, the rover continues collecting samples and exploring the Jezero Crater, proof that it has not only transformed Mars into an endurance track.
CERN and Mistral are finally preparing the Flare project, a basic model trained not on texts, but on billions of simulated proton-proton collisions. It must reconstruct the events recorded by the Large Hadron Collider, identify particles and detect atypical phenomena. The agreement remains a collaborative principle: Before pursuing new physics, researchers will need to measure the model’s uncertainties, compare its results with established methods, and verify that it hasn’t simply learned biases from its simulations.
