OpenAI has decided to temporarily slow down the scaling of its most advanced artificial intelligence models to strengthen its security mechanisms, a decision that reflects the industry’s priority to keep control systems at the same level as the capabilities of a new generation of increasingly autonomous models.
“As models gain greater capabilities, the risks associated with their development and internal testing also increase.”. Our supervision, alignment and security standards must stay ahead of these risks,” the company explained.
The decision comes after OpenAI announced in early August the suspension of certain tests of a not yet published model, called Astra, when “significant advances in agent programming and cybersecurity” were detected in internal evaluations.
Added to this are other episodes and tests recently disclosed by companies such as Anthropic and specialized organizations, which have once again brought to the fore the ability of the most advanced AI systems to execute complex tasks and operate with an increasing degree of autonomy.
These events reveal an evolution that goes beyond the improvement of the capabilities of the models: artificial intelligence is going from being fundamentally an assistance tool to becoming a technology capable of executing actions increasingly autonomously. This transformation forces organizations to rethink their traditional approaches to technology risk management, cybersecurity, and digital governance.
For Fernando de Águeda, CEO of Scalian Spainthis evolution should not be approached from an alarmist perspective, but from need to adapt risk control and management mechanisms to a new technological reality.
“Artificial intelligence does not introduce a completely new risk; it amplifies the speed, scale and complexity of risks that we already knew. That is why the conversation should not focus on the fear of the autonomy of AI, but on our ability to effectively monitor, audit and govern it.”
Greater capacity requiring new control mechanisms
The evolution of foundational models and intelligent agents is simultaneously increasing their capacity, speed of execution, and ability to operate at scale. For companies, this advance opens new opportunities for automation and productivity, but also introduces new challenges related to supervision, access control and risk management.
In this context, the adoption of artificial intelligence ceases to be a question of the future and becomes a business reality. The challenge is to ensure that deployed systems operate within strong frameworks of trust, transparency and accountability.
“The greater the decision-making capacity of a system, the greater must also be our ability to know what it is doing, why it is doing it and what limits it has”says de Águeda.
The expansion of artificial intelligence represents a relevant change in traditional security models. As happened with the Internet, cloud computing or the Internet of Things, the incorporation of new technological capabilities expands the possibilities of organizations, but also generates new risk vectors.
In the case of AI, the change is especially significant because organizations will have to manage ecosystems in which people, applications and intelligent agents interact with each other, in some cases with the ability to execute actions without direct human intervention.
Therefore, capabilities such as model observability, decision traceability, continuous auditing, rigorous identity and permission management, human supervision, and Zero Trust architectures are especially important.
“The challenge is no longer just about protecting an infrastructure against external threats.”. We have to be able to govern systems that can act, interact and make decisions at a speed much higher than that of traditional supervision processes,” explains the CEO of Scalian Spain.
Resilience as a new security paradigm
The growing autonomy of AI also forces it to evolve from fundamentally reactive security models towards digital resilience strategies, capable of continuously adapting to new scenarios and behaviors.
For organizations, this means incorporating security by design, establishing clear limits of action for agents, strengthening AI governance mechanisms, and maintaining permanent monitoring, auditing, and response capabilities.
In critical and highly regulated sectors, these mechanisms take on even greater importance, as organizations must be able to demonstrate not only that their systems work, but also that their decisions can be monitored, mapped and corrected.
“AI systems will become increasingly autonomous, and that evolution is irreversible. The real strategic issue is ensuring that every decision can be explained, mapped and corrected.. Trust in AI will depend less on what it is capable of doing and more on the quality of the control mechanisms we build around it.”
The incorporation of artificial intelligence into business processes should not be considered as a trade-off between innovation and security. For Scalian, both dimensions must evolve together and be part of the design of digital transformation strategies.
“The organizations that will be successful will not be those that slow down the adoption of AI, but those that are capable of combining innovation, governance and operational resilience from the first moment. The competitive advantage will not only be in having more capable systems, but in knowing how to use them in a safe, controlled and responsible way,” concludes Fernando de Águeda.