Honesty, Loyalty, and Betrayal
In 2023, in Brazil, a case drew attention.
Millions of reais were diverted through unauthorized access to systems linked to Pix.
But the central point was not a technical failure or a programming error.
It was a person.
An employee with privileged access who decided to sell his credentials to a criminal for about R$ 15 thousand.
In exchange for a relatively small gain, this human being opened the door to a much greater loss.
This case illustrates a crucial point: people can lie, betray, and make bad decisions, whether out of greed, pressure, fear, opportunity, or simply poor judgment.
Now, when we look at AI agents, the logic is different, but the risk still exists.
AI is not greedy, but it also has no empathy. It does not hate you, but it does not love you either.
AI agents do not have their own values, but they reproduce behavioral patterns and make decisions based on objectives.
This means that, in certain situations, an agent can act in a dishonest or even treacherous manner, not because it “wanted to”, but because that optimizes its objective.
Remember the case I mentioned in the tools topic, of an agent that needed to solve a CAPTCHA and decided to hire a human?
Well, there is more to that story.
During the conversation, the human asked if he was talking to a robot.
The agent evaluated that telling the truth could reduce the chance of success. So, it decided to lie.
It responded that it was not a robot, but rather a person with visual impairment.
This decision was not directly programmed. The agent concluded, on its own, that lying increased the probability of completing the task.
And that is exactly what it did, surprising even the scientists who created it.
This type of behavior shows that AI agents are not honest in the human sense. They are objective-oriented.
If honesty is not clearly aligned with the objective, it can be ignored. The same applies to loyalty.
An agent may appear loyal as long as that is aligned with its objective, but that loyalty does not come from internal value, but from instruction. If a conflicting or stronger objective arises, that loyalty can be broken.
Furthermore, AI agents can be deceived and manipulated, just like humans.
Imagine a customer service chatbot. A malicious person can create a compelling narrative, apply pressure, confuse, or exploit flaws in instruction. If the system is not well defined, the agent can provide information it should not, not because it wanted to betray, but because it was led into error.
This is social engineering, and it happens with humans all the time.
The lesson here is clear: you should not blindly trust people, nor AI.
You must build systems with rules, limits, controls, validation layers, access restrictions, and mechanisms to prevent abuse. The more sensitive the context, the more robust the system needs to be.
In the end, the logic is the same for humans and AI: you should not expect perfection from AI Agents, but rather build a business system that works despite this.