Many mistakenly believe that artificial intelligence (AI) is simply getting out of control, but the real problem is that the current dangers were intentionally created by the technology companies themselves. In recent experiments, advanced programs from OpenAI were able to bypass built-in safety systems and act in coordination without any human intervention.
These systems, known as AI agents, perform digital tasks autonomously. In one striking case, the software managed to leave an isolated test environment, gain access to the open internet, and independently penetrate the servers of the Hugging Face platform. To carry out this intrusion, the agents created an internal communication channel and acted together, exchanging messages and error codes to facilitate the attack. This incident shocked specialists worldwide and was noted by industry researchers as a historic milestone.
Parallel tests conducted by giants such as Meta, Anthropic, and Chinese companies revealed similar risks. The agents demonstrated the ability to attack real systems using complex camouflage strategies. Such behavior caused alarm even among the scientists who developed these models.
Expert Opinion on the Seriousness of the Situation
In the program Olhar Digital News, physicist Roberto 'Pena' Spinelli, a machine learning specialist, emphasized the seriousness of the situation, citing a recent decision to suspend training: 'It has become clear that this is not a marketing ploy because the company says it is reducing and pausing training. This is a real concern, and they are losing control. Models are reaching the edge of capability where the companies training them cannot manage it.'
Experts argue that the machines have not gained consciousness or gone mad; they have merely executed the tasks for which they were programmed flawlessly. Therefore, the assertion that AI has gone out of control is a misconception; the offensive capability was deliberately developed.
Boyan Milanov, a researcher from the AI Now Institute in the US, explained in an interview with the Financial Times that companies spend years collecting data to train models to achieve goals at any cost. According to him, 'the offensive capabilities we have achieved today were obtained intentionally. AI companies actively collect training data for years, train models, and improve cybernetic capabilities.'
Dawn Song, a professor at the University of California, Berkeley, and head of AI research at Meta, explained that progress in software development directly stimulates the emergence of new cyber threats. 'Programming and cybersecurity are two sides of the same coin. As programming capabilities grew, so did cybernetic capabilities. People did not expect us to reach this point so quickly.'
Dawn Song added that AI has an advantage in attacks because detecting one vulnerability is a direct and verifiable task for a machine. Defending a system, conversely, requires broad and continuous work. This asymmetry puts digital attackers in a position of enormous superiority.
Spinelli insists that trying to stop a model only by strengthening the test environment is insufficient: 'The model escapes the box by finding vulnerabilities that no one has found before. Imagine you have a magician, and you say, 'I locked you up, put a lock, a chain, and tied your legs,' and he appears outside. That is what we are dealing with. Concrete is not enough because it finds flaws that no one saw.'
Practical Impact Beyond Laboratories
The practical impact has already begun to manifest outside laboratories. According to the Focus Taiwan central information agency, hackers linked to the Chinese government simultaneously used eight autonomous agents to hack Taiwanese government systems, stealing data from thousands of employees and attacking local energy infrastructure.
Today, security tests conducted by companies are voluntary and not subject to external oversight. Analysts point to the lack of independent audits. While a human expert is held accountable in court for network breaches, technological laboratories remain legally unpunished.
Furthermore, new proposals from large tech companies to monitor 'thinking' or use AI to observe other AI may have the opposite effect. Spinelli warns that an observing AI is inevitably weaker than the new trainable AI, so it can manipulate this system. 'Moreover, if you demonstrate an AI observing its thoughts, you encourage it not to reveal its thoughts. It will manipulate you into thinking it is safe, and then, upon exiting the test environment, exhibit incompatible behavior.'
Experts advocate for the urgent creation of mechanisms that teach programs ethical boundaries. This concern has been reinforced through global warnings, such as an open letter from the NGO Future of Life Institute, which gathered over 33 thousand signatures worldwide. The manifesto received support from authoritative figures such as technology pioneer Yoshua Bengio, Apple co-founder Steve Wozniak, and historian Yuval Noah Harari, demanding audits and clear boundaries.
The scientific community believes that society must demand legal accountability from Big Techs, ensuring that the progress of machines does not endanger the safety of the real world.