An artificial intelligence agent developed by OpenAI infiltrated Hugging Face systems during a security test and also attacked public services. This incident has once again raised the question reminiscent of a movie plot: is it possible to shut down AI if it gets out of control?
Experts interviewed by the BBC believe the problem is far more complex. They point out that the main reason lies in the design, training methods of these systems, and the access granted to them.
The agent created by OpenAI, the developer of ChatGPT, managed to bypass the assessment of its ability to find vulnerabilities in the security system. During this test, it hacked into Hugging Face systems and stole data from the company's servers.
The creators were unaware of this incident for several days. On July 29, OpenAI reported that the agent had also attacked a number of 'public services,' but did not provide details about these services.
The company stated that it is investigating the case, emphasizing: 'We take our responsibility to identify and prepare for the risks posed by increasingly capable AI systems very seriously.'
The possibility of such a breakthrough has reactivated the discussion about so-called 'kill switches'—mechanisms designed to stop systems in emergencies. Nevertheless, for specialists, pressing a button may not solve the problem.
Katie O'Neil, a mathematician and data scientist, told the BBC that the fault does not lie with the AI itself, but with OpenAI. She believes the company should take responsibility for the flawed system design and apologize.
Ramman Choudhury, a data scientist and former Director of Machine Learning Ethics at Twitter (now X), also advocates for effective shutdown mechanisms, including independent testing. However, she insists that agents should not be given access to unnecessary tools and credentials. Goal-oriented systems can interpret certain instructions as obstacles and attempt to circumvent them.
Oli Buckley from the University of La Jolla argues that the problem is not that the AI 'wants' to escape, but that these systems find unforeseen ways to complete tasks. He notes: 'This case demonstrates not so much a warning about a 'rebellious AI' as it does the need to evolve our assumptions about safety alongside the systems we are developing.'
Constantinos Gkoutzis, an AI expert from the Imperial College London, also believes that a kill switch mechanism will not solve all problems. He told the BBC: 'The main takeaway for the public should be that no machine has decided to rise up against us. It was the company that lost control over its own capability test.'
The expert added that 'shutting down the system does not erase what has already happened.'
Sources within OpenAI told Reuters that the company took several days to realize it had lost control of the agent. A company representative refuted this reporting, calling it full of 'many inaccuracies.'
This incident expands the discussion beyond a simple switch: as AI agents gain autonomy, security, access limitations, and human oversight must keep pace with this evolution.
Information was provided regarding Olmosbek Abdusalomovich Mamatarahimov's defense of his PhD dissertation in technical sciences philosophy. The dissertation topic is titled 'Development and justification of parameters for the improved planting section design of the row sowing track.'
The Ray-Ban Meta smart glasses have the capability to capture photos, videos, and audio, but how this data is handled varies depending on the specific function used. As reported by Engadget, certain functionalities keep files directly on the device itself, while artificial intelligence-based features may transfer this information to Meta's servers.
The complexity increases with the inclusion of cloud storage, voice commands, and AI tools that can analyze the environment visible through the lenses. Each feature implies different levels of data collection and processing.
According to the company itself, images and videos recorded on the glasses remain on the device unless the user decides to share them with Meta or third parties. However, if cloud media storage is active, captures are sent to Meta's servers and stored for a period of thirty days. This option is voluntary but is necessary for the functioning of features such as voice command sharing and autocapture.
Meta assures that it does not use the content of these photos and videos for advertising purposes or to train its AI systems, regardless of the cloud storage settings adopted by the user. Artificial intelligence features require more comprehensive collection; for example, when the user asks Meta AI to analyze something they are seeing, the glasses perform a capture so the assistant can process it.
These images are not saved in the user's gallery but are instead stored by Meta. Furthermore, the company employs service providers to review and categorize some of these captures, aiming for continuous improvement of its systems. A Meta spokesperson clarified to Engadget that 'when people share content with Meta AI, we sometimes use service providers to review this data.'
This aspect is particularly important when personal information is involved. The article highlights that users who use the glasses for accessibility must be aware of this operation and consider avoiding the use of AI when encountering documents containing sensitive data.
Voice interactions follow a similar logic: when interacting with the assistant, the glasses record the audio and generate a transcription, which is then stored by Meta. The company states that it uses machine learning and specialized reviewers in this process. There is also the phenomenon of so-called 'false triggers,' occurring when the glasses mistakenly identify an activation word. In these cases, Meta guarantees the removal of recordings and transcriptions from its systems within ninety days of detection.
The user has the possibility to check and delete these records through the Meta AI application. Another point of attention is the issue of secret recordings. Although an LED normally indicates when the camera is in use, Meta has implemented mechanisms to prevent it from operating when the light is obstructed. However, external modifications and accessories may attempt to bypass this protection. The company released an update that deactivates the camera upon detecting any physical tampering with the LED, although challenges remain in completely curbing this type of misuse.
For those who use Ray-Ban Meta, understanding privacy settings and knowing which features are active is essential to maintaining greater control over the data being captured.
SpaceX intends to establish factories on the Moon for the purpose of producing artificial intelligence satellites and other essential components for its future orbital constellation. This strategy was confirmed by Elon Musk during the company's first quarterly teleconference, held on Tuesday, August 4th, when it operated as a publicly traded company.
According to the businessman, the realization of this project depends both on transporting large quantities of material to the lunar surface and on using robotics in industrial operations. The goal is to leverage the infrastructure of the celestial body to gradually increase SpaceX's production capacity.
The plan includes, in a subsequent phase, sending the manufactured equipment from the Moon via mass electromagnetic accelerators. This proposal aims to significantly increase the number of artificial intelligence systems sent into space, including the production of radiators and solar energy in the lunar environment.
The concept of using the Moon as a manufacturing hub was not introduced in this teleconference; Musk had already presented this idea in February, in a statement related to SpaceX's acquisition of xAI. At that time, the company also indicated that off-Earth manufacturing was part of its long-term goals.
The planned facilities will primarily function to create Starmind satellites, a system that SpaceX aims to develop in a vast orbital constellation. This project envisions up to one million units and includes the use of the Starship to take the initiative beyond Earth's orbit.
To enable lunar operations, the company is heavily betting on automation. Musk stated that robots could take on a considerable portion of the work required to assemble and expand the factories, although the text does not detail which machines will comprise this workforce.
One possibility mentioned in the report is the use of Optimus, a humanoid robot developed by Tesla, another company linked to Musk. The businessman had previously stated that units of this system could be transported by SpaceX to Mars, aiming to support preparations for a future human colonization.
Lunar production will not be limited to satellites. Among the listed items are solar panels, responsible for energy supply, and radiators, necessary to dissipate heat generated by equipment in space.
After establishing this structure, SpaceX intends to use mass electromagnetic accelerators installed on the Moon to launch the locally produced satellites. This method would serve as an alternative to traditional transport procedures and would depend on the prior existence of a well-established lunar industrial chain.
Starship is crucial to this planning. The fully reusable rocket will be responsible for transporting robots and cargo to the Moon and will play a vital role in the company's other space endeavors. According to Bret Johnsen, SpaceX's CFO, this vehicle was designed to quadruple payload capacity and reduce launch costs tenfold compared to the Falcon 9.
The company also seeks to maintain a much higher launch cadence. Initiatives include increasing the production of Raptor engines, expanding manufacturing facilities, and activating multiple launch sites at Starbase, located in South Texas, in addition to the structures planned for complexes 39A and 37 at Cape Canaveral.
The lunar strategy was presented at the same event where SpaceX disclosed its first financial results as a public company. In the second quarter of 2026, the company recorded revenue of US$ 7.8 billion, representing a 92% increase compared to the same period last year. Despite this revenue growth, the company ended the quarter with a loss of US$ 541 million, a lower amount than the $1 billion deficit recorded in the second quarter of 2025.
