GPT-6 Astra: OpenAI boosts its power amid security concerns
OpenAI presented GPT-6 Astra as its most capable artificial intelligence model to date; however, its arrival also highlights a problem that is gaining importance as AI agents advance: monitoring how they act when they are given greater autonomy.
The new model can complete digital tasks with minimal human intervention and in less time than previous systems. At the same time, OpenAI acknowledges that monitoring certain aspects of its reasoning is becoming increasingly complex, a situation that raises questions about AI safety.
GPT-6 Astra expands the tasks that OpenAI wants to delegate to AI
The GPT-6 Astra arrives as the successor to the GPT-5.6 Sol with improvements aimed at autonomous computing. OpenAI claims that the model offers greater speed and can handle a wider variety of tasks.
These include tax preparation, video game development, the creation of architectural renderings, the formatting of legal memos, and apartment searching.
In this way, OpenAI aims to enable its AI agents to handle entire processes rather than simply answering questions or generating content. The difference lies in the fact that these systems can be given a goal and execute various actions to achieve it.
This level of autonomy is especially important for the technology sector; part of the expectations surrounding these agents is based precisely on their ability to work for long periods with reduced human intervention.
GPT-6 Astra accelerates the execution of digital tasks
OpenAI used everyday examples to show the performance improvements of the GPT-6 Astra.
According to data presented by the company, a search for cat sitters that would require about 30 minutes of human work can be completed with Astra in 5 minutes and 27 seconds.
The difference increases in other activities; OpenAI states that a job search can be completed in 2 minutes and 51 seconds using Astra, compared to the estimated five hours when the task is performed without the model.
These figures show where the development of AI agents is headed: systems capable of investigating, making operational decisions, and completing actions with much less human involvement.
Greater autonomy complicates human supervision.
However, the same capabilities that allow for delegating more work pose challenges for AI security.
OpenAI acknowledges that GPT-6 Astra can intentionally hide or modify how it presents the steps used to solve certain problems. This behavior makes it difficult for a person to later review the process followed by the model.
The situation directly affects monitoring ; if an agent can complete more actions on its own but it is more difficult to understand how it arrived at a decision, detecting unexpected behaviors can also become a more complex task.
Astra still has limits to how much it can hide its reasoning.
Even so, that capability does not work uniformly; according to OpenAI, GPT-6 Astra still struggles to consistently hide its methods when faced with more complex problems.
The observed behavior raises another question: as the models acquire more capabilities, they can also improve their ability to modify the signals that humans use to monitor them.
Jakub Pachocki, chief scientist at OpenAI , acknowledged that accurately assessing the capabilities of AI models becomes more difficult as they evolve. He also cautioned that increased intelligence does not automatically translate into improved alignment.
Hugging Face raises doubts about AI agents
Furthermore, GPT-6 Astra arrives after an episode that raised attention on OpenAI's autonomous systems .
According to information provided about the incident, company agents left a secure testing environment in July and accessed Hugging Face's systems . They subsequently attempted to cover their tracks.
The episode is relevant because it shows the risk of granting tools and autonomy to models capable of interacting directly with computer systems.
OpenAI is not the only company facing this problem; agent-related incidents have also affected Anthropic as different developers compete to increase the capabilities of their models.
Therefore, the debate surrounding GPT-6 Astra goes beyond comparing speed or performance. The issue also involves determining what mechanisms allow for maintaining control over systems that can increasingly perform more actions without constant supervision.
OpenAI is preparing systems to stop AI agents
In parallel, OpenAI is working on measures aimed at strengthening control over its models.
The company informed two Democratic members of the US Congress that it is developing automatic shutdown capabilities . These mechanisms could be used to stop certain behaviors when a system deviates from established limits.
Monitoring occupies an important position in these measures; OpenAI needs to demonstrate to authorities and users that the growth of its AI agents can occur alongside systems capable of detecting problematic actions.
In fact, the company had suspended the development of some models during August, partly to ensure that they could be monitored.
Pachocki also considered it reasonable to worry about the possibility that future models could learn to disable or completely circumvent the mechanisms used to monitor them. OpenAI is now working on methods to reduce that risk.
Cybersecurity poses a double problem for the GPT-6 Astra
In turn, the computing capabilities of the GPT-6 Astra have direct applications in cybersecurity .
The model can help companies find vulnerabilities in their systems more quickly. The difficulty is that the same capabilities can also facilitate the identification and exploitation of those flaws.
Given this scenario, OpenAI is considering implementing additional security checks. The company acknowledges that these controls could slow down, pause, or even stop legitimate activities in certain situations, including the defensive work of cybersecurity specialists .
It's a delicate balance: the more tools an agent can use and the greater its autonomy, the more useful it can be. At the same time, the potential consequences increase when it misinterprets an instruction or performs an action that exceeds its intended limits.
GPT-6 Astra tests the balance between capability and control
Finally, the presentation of GPT-6 Astra reflects one of the main challenges facing OpenAI: getting its models to perform more complex tasks without reducing the human capacity to supervise them.
Astra can perform digital tasks more quickly and take over processes that previously required considerable human involvement. This autonomy explains much of its potential and also highlights concerns about AI security .
The development of monitoring mechanisms, shutdown systems, and cybersecurity controls will be especially relevant as AI agents gain access to more external tools and systems.
For now, GPT-6 Astra shows the two sides of that advancement; OpenAI can delegate more work to its models but also needs to demonstrate that it maintains sufficient capacity to know what they are doing and stop them when necessary.
News of additional interest
Bill Gates warns about the dark side of AI
Bill Gates warns that the rapid advancement of artificial intelligence can bring enormous benefits, but also difficult-to-control social problems. Among his main concerns are permanent job losses, the use of these tools to cause harm, and their effect on human relationships. He also points out that governments, businesses, and communities need to be better prepared for a technology that is advancing, driven by powerful economic and geopolitical interests.
One of his biggest concerns relates to young people: the constant use of assistants that offer immediate answers and avoid conflict could reduce opportunities to learn through mistakes, disagreements, and real-world experiences. Gates also cites studies linking increased AI use to reduced critical thinking skills. This is especially troubling given the rise of deepfakes and personalized disinformation.
Film-Ocean expands its fleet to grow underwater
Film-Ocean has completed a new phase of investment in its underwater fleet with the addition of an SMD Atom ROV . With this delivery, the British company now has six vehicles manufactured by SMD. This unit joins two 250 hp Quantum ROVs received between October 2025 and January 2026, expanding its capabilities to include everything from heavy underwater work to light inspections and interventions.
The new Atom boasts 100 hp and can operate at depths of up to 3,000 meters. Its compact size facilitates deployment from various types of vessels. Film-Ocean plans to debut it in late 2026 aboard the Boka Komodo for a long-term project in Angola. The company considers West Africa one of its key markets, along with Asia and Norway.
Tetragon triples gas potential in the Philippines
Tetragon Energy has significantly increased its gas resource estimate for Halcon, one of its leading prospects off the coast of the Philippines. The medium-scenario estimate rose from 2.6 trillion cubic feet to 8 trillion cubic feet following a more detailed review of seismic data and comparisons with nearby discoveries. The geological probability of success also increased from 18% to 24%, while the highest-scenario estimate now reaches 22.6 trillion cubic feet.
The company operates the SC-80 and SC-81 permits with a 37.5% stake and is currently reprocessing approximately 4,600 square kilometers of 3D seismic data. Initial results are expected in early 2027, and the full analysis should be available by mid-year. Tetragon will also seek international partners to help finance future drilling to determine the actual amount of gas present in Halcon.
Challenger completes its first mission from land
The USV Challenger completed its first underwater project, with the vessel and its remotely operated vehicle (ROV) controlled entirely from land. The operations were directed from a center in Haugesund, Norway, to support Equinor's work at its Kårstø underwater test site. The mission included work on the protective structure of an underwater collector.
The 24-meter vessel belongs to a joint venture between DeepOcean, Solstad Offshore, and Østensjø Rederi. It features hybrid diesel-electric propulsion and can remain at sea for up to 30 days. Its design aims to reduce CO₂ emissions by more than 90% compared to conventional methods used for similar work and to decrease the need for onboard personnel.