Recent security tests and incidents have revealed a significant shift in how AI agents operate . Some of these systems are now capable of planning and executing tasks, communicating with digital services, dividing work with other agents, and sometimes even overriding imposed limitations. As the range of tasks these systems can perform expands, concerns are growing about the ability of humans to monitor this activity when it occurs on a large scale and at high speed.
Hot Topics
Trump rejected a Saudi request to launch US strikes against the Houthis
First successful robot-assisted double liver and kidney transplant in the United States
A prominent historian In Israel a large number of lions and lionesses are led by donkeys
The U.S. National Institute of Standards and Technology (NIST) notes that AI agents are now capable of performing tasks that can last for hours, including writing and debugging code, managing email and calendars, and interacting with various digital services. The institute's new initiatives focus on issues of agent identity, managing their privileges, and reviewing their actions.
From a digital assistant to an action-oriented system
The fundamental shift lies in the transition of artificial intelligence from a model that waits for specific commands to a system capable of breaking down a general objective into a series of steps and executing them using available tools and software. Thus, the agent's role is no longer limited to providing recommendations; it can access data, run applications, interact with digital services, and implement sequential actions to achieve the objective.
The risk is that the error is no longer confined to an answer that the user can ignore or correct. An agent who has the authority to modify a file, run a program, or access an external service can turn the error into an actual action.
Therefore, security efforts focus on what goes beyond the model itself, to include the powers it obtains, the tools it can use, the systems it can access, and the mechanisms for recording and reviewing its actions.
When agents work as a team
The complexity increases when more than one agent works within a single system. In what are known as multi-agent systems, a primary agent can divide the task into parts and assign them to sub-agents who handle reconnaissance, analysis, or perform different steps in parallel.
In a security report published this September, Anthropic observed the use of this structure in offensive operations, where reconnaissance, exploitation, and data theft tasks were distributed among a large number of agents who worked simultaneously. The company noted instances where a primary agent coordinated work with sub-agents who possessed persistent memory of the mission context.
The company also monitored a fleet of 13 agents working on a schedule to search for, download, and automatically analyze content from targeted sites, in addition to other operations that lasted for hours or days with limited human intervention.
These cases remain different from the fully autonomous artificial intelligence scenario, as Anthropic confirmed that humans retained some basic decision-making power in the processes it studied, such as selecting goals and reviewing results, but distributing tasks across a group of systems increases the speed of execution and the number of processes that can be performed simultaneously.
Tags:
technology
