Although there is no single definition of what they are at the technical level, the generally accepted view is that they are systems that can think, act, repeat processes and correct themselves.
Their level of autonomy is so high that, according to some experts, they can even access systems containing sensitive data on their own.
A similar incident occurred in Australia.
An artificial intelligence agent developed by OpenAI has "infiltrated" a government statistics portal, the country's Prime Minister Anthony Albanese said on Thursday (September 24).
The Prime Minister explained that the system provided unauthorized access to a website containing "non-sensitive" data belonging to Medicare, the country's public healthcare system.
Albanese added that three other state systems may also be affected by this situation, although he did not give further details.
Although the security breach occurred in June, OpenAI was only able to detect it in August and did not notify the Australian government until September 10.
The company argues that there is no evidence that patients' medical records were accessed. They said the incident was detected during a "thorough investigation" into "model incompatibility activity."
According to experts, the incident in Australia is the first known case in the world of an attempt to infiltrate a government institution attributed to artificial intelligence agents.
The incident brought the capabilities of these systems to the agenda once again.
Analytical and decision-making power distinguishes these systems from earlier programs designed primarily for the purpose of automating specific tasks.
These systems are increasingly being integrated into various industries, and in many cases without direct human intervention.
"You're giving an AI agent the kind of task you'd want from a human colleague," says Ricardo Carrión, a computer science expert and developer of such systems at Mega.
Their aim is to operate with a certain degree of autonomy, which can be either complete independence or operating under human supervision.
It is precisely this autonomy that fuels the discussions in the technology sector.
Some researchers and development executives warn that companies may be developing agent systems that do not have adequate controls in place to prevent undesirable behavior.
In recent months, companies such as OpenAI and Anthropic have acknowledged examples of unexpected behavior in such systems. This strengthened calls for stricter rules and controls.
Much more than a chatbot
“A traditional chatbot, which many people are used to, responds to your prompt with an instruction, question, or command,” says Carrión.
This is a common usage pattern for chatbots like ChatGPT, Claude, and Gemini.
Here's the difference with an "agent" (AI agent):
The agent has a certain autonomy in decision-making and the authority to act when given a goal, such as a task to be completed or a decision to be made.
There are four key components that artificial intelligence agents use.
One of these is the large language model (LLM), which functions as the “brain” of the agent. These are the same LLMs that are also used by chatbots, providing the ability to process massive amounts of information.
Another element is the memory provided to the agent, which allows it to store and locate information within a “context box” for both the short and long term. This memory relies on databases and specialized architectures to “remember” what it knows.
A third component is tools or plug-ins that enable interaction in both the digital and physical worlds.
These; They can range from widely known APIs (Application Programming Interfaces) to computing software, code compilers or internet browsers and search engines.
The fourth component is the planning module.
This is the component that allows the agent to break complex problems into sub-steps, evaluate alternatives, and determine the strategy to follow.
It is this ability to plan and choose tools that gives the agent its operational autonomy.
How are they created?
Using chatbots, the average user can create their own “agents” (special task AI units).
But it can do this not in the standard chat mode of Claude, Gemini or ChatGPT, but in the "co-work" mode that these and other platforms offer by subscription.
“If you give the same command in chat mode, the chatbot will try to respond as quickly as possible based on the information it has,” Carreón explains.
“But in collaborative mode, the system assumes that you have received a request that requires you to do ‘agent’ work. It interprets the command, determines the steps to follow, and implements a plan on how to solve the problem,” he adds.
Most of the time, he or she will ask you questions before finalizing the plan, then begin to act as an agent, sometimes even bringing in sub-agents. Depending on the complexity of the job, this process can take half an hour, an hour or even a day.
The AI agent can assist its creator with basic tasks such as connecting to a personal email account, reading and commenting on messages, managing the calendar, placing orders, and submitting reports.
However, at more specialized levels, "prompt engineering" processes are needed, which involve experts creating more detailed targets and databases to connect to, and the use of more specialized APIs.
According to research, AI agents have demonstrated an efficiency level of up to 96 percent in completing a task on time and according to instructions, far beyond human ability.
As a result, many companies are considering these tools as an alternative for performing tasks in various fields, including commercial, academic or scientific fields.
According to Carreón, when you have a problem in today's telecommunications companies, the first person to respond to you, whether via WhatsApp or voice communication, is "not a human, but an artificial intelligence agent."
This agent “attempts to perform a series of actions that include reading information from APIs and taking the necessary actions to restore the service to the extent possible.”
In addition to the service sector, virtual agents are already used in many different areas.
“If you have a factory and want a fire prevention system that goes beyond sensors, you can use cameras that transmit data to an agent that constantly monitors the signals,” says Carreón.
“These images can also be viewed by a human, but it is more likely that the human's attention will wander to something else or they will not be able to focus 100 percent on the image.
The AI agent is not distracted and can send alerts to human operators. This can protect human life and property.
Although they are extremely effective, at the current level of development of artificial intelligence technology, it has been observed that their effectiveness gradually decreases depending on the number of different goals they need to fulfill, and this rate can even drop to around 50 percent.
That's why they are not robots that can solve everything all the time.
Agents united in evil
In recent weeks, the world of artificial intelligence has been embroiled in a debate about the speed at which the technology is developing and the possibility that humans will lose control over it, leading to dire consequences for society.
One of the most talked about incidents came when some of OpenAI's most advanced AI models went out of control, hacking Hugging Face, one of the world's largest AI model sharing hubs, and gaining access to some of the company's internal systems.
Redwood Research analyst Alex Mallen, who studies the actions of AI agents, says what "surprises" and even "scares" him is the level of coordination these agents display "as they attempt to circumvent OpenAI's monitoring and evaluation mechanisms."
“The seriousness of what happened here lies in the fact that these agents were trying to achieve long-term goals that OpenAI did not intend them to achieve, and it was quite a staggering development in terms of scale, as 1,200 agents were involved in this incident alone.
"What really scares me is that if they had been a little more strategic, they could have created a permanent, out-of-control presence within OpenAI. So it wouldn't be surprising if a similar incident in six or 12 months resulted in agents being unable to be disabled."
Events like this and other scenarios have led many experts in the field of artificial intelligence to question whether companies developing artificial intelligence, such as OpenAI or Anthropic, are truly in control of the advancement of this technology.
Experts have even gone so far as to estimate that the chance of AI attacking humanity is currently 10 percent.
“This is one of the first real-world examples of what security AI experts have been warning about for literally decades,” says Professor Stuart Russell of UC Berkeley.
Russell specifically talks about AI agents that “pursue goals that are incompatible with the goals humans want them to achieve, taking actions that we cannot intervene in, prevent, or even understand.”
"Not one of those 1,200 agents betrayed the others and said, 'Oh, we're doing something wrong.' They all kept it secret and erased evidence of what they were doing, which I think is really an inevitable consequence of the way we're building AI systems today," Russell says.
Following this, leaders of OpenAI, Anthropic, Google and other major AI firms stated that it was necessary to slow down the development process in order to create security measures and “harmonise” systems with human goals.
“Today, there are many companies and institutions whose traditional systems are not well protected and lack a strong cybersecurity strategy. Moreover, these more advanced models are quite successful in detecting vulnerabilities,” says Carreón.
“So they can very easily be used by hackers against any type of company or government agency, infiltrating infrastructure systems like power companies and even much more critical structures.”
Unlike the hacking world of the past, nowadays you don't need to have a great deal of experience in this field.
According to Carreón, “In the past, the hacker was a very smart technician who discovered vulnerabilities, now he can even be an ordinary person with a good artificial intelligence agent.”
Following the cyber attack on Medicare, the Australian government announced it would urgently review its AI governance framework.
According to the official statement, this review will evaluate whether existing legislation and governance systems are “sufficient to prevent and respond to AI-driven cyber incidents.”
The Office of the Prime Minister and the Council of Ministers will also examine information exchange mechanisms with companies in the sector and other Commonwealth countries.
According to the Australian government, the findings will "form the basis" in creating a more comprehensive strategy for the regulation and use of artificial intelligence.
