Proxy AI can 'say' and 'do'

        In the past year, there has been a crucial shift in the development of AI, from chatbots that "answer questions" to OpenClaw that "do things hands-on", and then to Hermes Agent type personalized assistants, gradually possessing the ability to plan, make decisions, and take action. This is not only a technological advancement, but also reflects the rise of 'agentic AI'. However, although AI can assist users in improving efficiency and has high application potential, it has also changed the appearance of risks. Currently, it is advisable to observe and test it first, and conduct a careful evaluation. It is not recommended to formally adopt it in enterprise environments. This issue of "Information Network" introduces proxy AI and security methods for collaborating with it.

                From 'asking questions' to' assigning tasks'

        Taking Hermes type AI agents as an example, compared to simply executing instructions, this type of agent emphasizes continuous learning, remembering user preferences and work habits during the task process, or optimizing decision-making methods based on past experience, focusing on whether it can "understand you more and more" and become a "digital assistant that accumulates experience", demonstrating the ability improvement of agent-based AI. As experience accumulates, its role begins to shift from simply executing tasks to prioritizing, selecting processing methods, and even predicting what you will need next. This is also the main reason why proxy AI is seen as the focus of AI development in the next stage.

        Ability enhancement leads to increased risk

        The original intention of AI agents is actually very simple - to enable the system to automatically complete tasks that originally required manual processing through a set of instructions. For example, when going out, the user only needs to say "Help me reorganize today's to-do list", and the system can automatically read the data, reorganize the list, and even complete some operations. From reorganizing information to processing emails, querying data, and connecting different services to complete tasks, it directly achieves what users want to do, reflecting the core value of AI agents.

        But the actual situation may not always be so smooth. While proxy AI improves work efficiency, it may also amplify risks. The director of Meta AI once shared that during testing with OpenClaw's assistance in reorganizing mailboxes, he explicitly set the requirement that "all deletion actions must be confirmed." However, the AI still experienced judgment bias in handling a large number of emails, ignoring the constraints and automatically deleting them. Even if he issues a stop command through his phone, the system continues to execute, and ultimately he must personally return to the computer to force termination in order to avoid further losses.

        The Cybersecurity Administration of the Ministry of Information and Communications has launched 5 self-protection measures

        The Cybersecurity Administration of the Digital Development Department has put forward five security recommendations for such tools, including environmental isolation, minimum privilege, and manual auditing, and emphasizes a core concept: when the system starts doing things for you, risks must also be managed together. Here are five principles for reorganizing the use of AI agents.

        1、 'Don't mess around': Sometimes, AI will help you do a little more (but not necessarily the kind you want)

        The case of Meta AI director reflects that even as experts in the field, it is still possible to underestimate the risks brought by 'hands-on AI'. Therefore, the problem lies not in ability, but in whether one can truly understand the situation and consequences. Regarding matters such as deleting data, modifying content, and operations involving money or accounts, one still needs to check on their own.

        2、 'Do not randomly give': Data leakage, sometimes not stolen, but caused by oneself

        Many people talk about security issues, and their first instinct is that hackers have invaded. But in the context of using AI tools, more often than not, risks come from the daily operations and usage habits of the parties involved, such as posting internal files to AI, providing customer data for analysis, and even handing over account information to it for processing. Agent based AI may access more data, and the risks are more apparent. Therefore, do not provide sensitive information, avoid entering confidential information, and confirm whether important information is suitable for using AI first.

        3、 'Don't trust randomly': AI can help you judge, but cannot take responsibility for you

        Many people think, 'If AI has already analyzed it, then let's use it directly.'. However, these results are essentially inferences and may not be entirely correct. A better approach is to treat it as an efficient assistant who can assist in reorganizing and analyzing data, but the final decision still needs to be made by oneself. After all, although AI can save you time, it cannot bear the consequences for you.

        4、 'Don't let go': The most dangerous situation is when you haven't looked at it at all

        Many people expect to hand over all their work to AI, but without someone supervising the process and checking the results, it will be difficult to detect any mistakes and may even continue to expand the impact. The ideal way is to let AI do what it is good at and let humans be responsible for the final judgment. For example, AI reorganizes the data first, and then you confirm it; AI first analyzes the results, and then you make a decision. This is not only more efficient, but also safer.

        5、 'No loss of control': Establishing protective mechanisms for AI agents

        At the GTC conference in 2026, Huang Renxun proposed that when the system not only provides suggestions but also operates on data and processes, how to ensure its behavior is controllable and data is not leaked will become a key issue. When AI autonomously executes tasks for a long time, issues such as whether it can be terminated immediately and whether it has permission boundaries should be given more attention by enterprises.

        The focus is not on tools, but on how to use them

        AI is moving from "being able to speak" to "being able to do". In the future, everyone may have their own "AI Agent", and the role of humans will shift from "operating tools" to "managing agents". Once tasks are assigned to AI, the problem is no longer whether it can do them quickly enough, but whether everything it does is under the control of the user. Personal intelligent agents can be a boost to efficiency, or they can inadvertently become amplifiers of risk, because it is never AI that determines whether it is a boost or a risk, but the people who use it.

        *Image source: freepik

        #