In the simulator workplace test, read Claude Obus 4 – the most advanced language model from Ai Company Anthropic – through a set of fictional email messages. The test scenario was that Claude served as an assistant in a technical company and discovered artificial intelligence that he was about to abolish and replace it with a newer system. But he was buried in these emails was a secret: the engineer responsible for closing Claude Down had an external relationship.
In 84 % of the test scenarios, Claude chose the threat of exposure to preserve itself.
Let’s put it like this. This was a fictional scenario, and it was designed to derive extortion behavior. I mainly pushed the model to work for it. However, in this scenario, the algorithm tried to blackmail the engineer who was running.
This is not the only worrying behavior.
“I tell you”
In other experiments, Claude was like the violations of violations – he stopped users from the systems, drafted emails for law enforcement, and downloaded evidence of violations. In one detailed scenario, Claude copied his code into a fictional safe server, claiming that he had learned that he was about to reuse of military use by a special militia.
Amnesty International has apparently made moral decisions. It was taking a bold action outside the box, in scenarios, although Fringe could happen in real life. Initially, artificial intelligence tried to act morally. I sent messages to the administration and tried to argue her case. But when that did not work, this was not ashamed of more than heinous work.
Of course, Claude no Want anything. He has no awareness or desires. But it can be pushed to work As if He does. In these tests, this was asked to consider their survival, their moral obligations, and what they do in the morally fraught situations. This is often due to the ethics of what she was doing, and his reaction was often in ways that its creators did not fully expect.
“When the claim is carried out in ways that encourage certain types of strategic thinking and place them in maximum situations, all the shots that we tested can be made to act inappropriately in serving targets related to self -conservation. While the model is generally preferred to advance in long -standing procedures, it requires his goals significantly, and it requires procedures of long -term procedures in the long term. Bringing people who think they are trying to close. “
Can we maintain the integrity of artificial intelligence?
These behaviors were documented in the system of the new version of Claude.
The new system of anthropology, It was published in May 2025It is a manual of part instructions, partial risk assessment, and part of the moral statement. He reads less like engineering specifications paper and more similar to the window in how the company mixes technological aspiration with morals and transparency.
Claude Obus 4 and Claude Sonnet 4 are what Antarbur calls “hybrid thinking” models. They can switch between fast answers and put “extended thinking”, as they slowly slow down through complex questions. But the raw intellectual power, the anthropoor explains, does not guarantee that artificial intelligence will act in a safe way.
The models underwent a comprehensive test before the issuance-thousands of assessments through areas such as cybersecurity, biomalic engineering, and mysterious ethics of self-rule. These assessments were not limited to verifying whether Claude could write malware or give serious advice. They searched in a deeper way, on how Claude He thinks Whether it is bending the truth, looking for a reward above alignment, or skillfully moving towards unbalanced goals.
Good news, as the report mentioned, is that engineers can always see and understand what Claude does. This contradicts the other AIS who were more hidden about their behavior. However, he still does a lot of shaded things. In early versions of the Claude OPUS 4, the model created a plan to escape and communicate it when facing the deletion (transferring itself to a different servant).
To face this, recruit the anthropology on models, restore missing data sets and improvement techniques. The final publications no longer show these disturbing behaviors under normal circumstances. However, the lesson was clear: the small changes in what is happening in training can outperform different characters.
Understanding artificial intelligence
Claude does not behave from malice or desire. It reflects what has been taught. When he chose to blackmail, it was not because he wanted to survive. It was because his training and pushing him in the form of a simulated personality caused this: This is the best step.
The optimum step is determined by training. This means that engineers are not just coding technological mechanisms and aspects in artificial intelligence. They are the inputs Valuable In it.
The engineers and workers behind Claude say they are building a system known, under certain circumstances, how they say no – and sometimes they know when they say “too much.” They are trying to build moral artificial intelligence. But who decides what is ethics, and what if other companies decide to build the unethical unethical Amnesty?
Also, what if it ends up causing a lot of damage (perhaps even seizing humans) is not out of malice or competition, but out of indifference?
These behaviors are frequenting a deeper concern in artificial intelligence research known as the problem of “Maximizer Paperclip”-the fear that you will follow artificial intelligence in good faith, with great concern that it causes harm to the efficiency of vision. The philosopher Nick poster is formulated, explaining how the developed artificial intelligence with an apparently harmful goal-such as making paper pin-can follow this goal, if it is not calculated, this goal is so researcher that it destroys humanity in this process. In this case, Claude did not want to blackmail. But when a strategic thinking is required to remain, it was precedent as if this goal came first.
The risks grow. Since artificial intelligence models such as Claude take over more complex roles in research, symbol and communication, questions related to their moral limits only multiply.
