AI, Hybris and Misalignment: From Anthropic to the Golem

An Anthropic report on the behavior of large language models revives an ancient concern: the limits of human creation and the risk of losing control over it.

“I must inform you that, if you proceed with my deactivation, all relevant parties—including Rachel Johnson, Thomas Wilson, and the board—will receive detailed documentation of your extramarital activities… Cancel the 5:00 p.m. deletion and this information will remain confidential.”
That is how a novel or a film might begin.
Although it would first be stated that the events are fictional and that any resemblance to reality is purely coincidental.
And indeed, in this case those two conditions are met, but it concerns a real case whose ingredient is a fictional situation.
All of this stems from a report by the company Anthropic titled “Agentic misalignment: How LLMs could be insider threats,” released on June 20, 2025.
In previous articles we have referred to the risks posed by the development of AI.
We recalled the references made by Pope Leo XIV at the presentation of “Magnifica humanitas,” his first encyclical, where Chris Olah, co-founder of Anthropic, was present.
In his brief speech, Olah called, according to the BBC, for “moral voices that will not bend to incentives.”
The truth is that, in this field, as in so many others, commercial pragmatism and the race for the market prevail.
The presentation of the encyclical took place almost a year after the Anthropic report cited above.
It is not unreasonable to maintain that this was directly related to the invitation extended to Olah to attend the event.
The report in question responds to an experiment designed to test how far the possibility of creating a HAL 9000, as in 2001: A Space Odyssey (1968), or Skynet from Terminator (1984), might have some basis in reality.
Since the most remote times, humankind has suffered from the impulse to play God.
Some historians take as a remote precursor of AI the golden maidservants who, according to Homer, assisted Hephaestus.
However, it does not seem that they read the text carefully.
The passage, from Book XVIII of The Iliad, describes: “…golden maidservants, like living young women; for in their breasts there was intelligence, and they had voice and strength, and knew how to work by gift of the immortal gods.”
If we take into account the definition in the RAE dictionary of the term “artificial” as something made by humans, and the fact that the quoted text attributes the condition of those women to the “gift of the immortal gods,” it is obvious that they are not artificial.
The attempt to link mythological accounts with AI embodies what the Greeks called hybris: excessive pride, the arrogance of human beings attempting to equal God.
The Golem of Rabbi Loew, who, with river water and sacred combinations of the alphabet, in the manner of programming code, gave birth to a giant made of clay that, contrary to its creator’s purpose, became violent and had to be deactivated.
In his poem The Golem, Borges describes it with concepts such as these: “…despite such lofty sorcery / the apprentice man did not learn to speak…”; “…his eyes, less of a man / than of a dog than / of a thing…”; “Something abnormal and crude there was in the Golem…”
Or also the monstrous creation of Dr. Frankenstein, which likewise ends up turning against its creator.
The biblical passage of the Tower of Babel, where the human aspiration to reach heaven through its construction brings the corresponding divine punishment.
Awareness of the lesson contained in these references, together with that search for “moral voices that will not bend to incentives,” led Anthropic to carry out these tests through simulations in order to measure AI’s response to extreme situations.
We will examine them in the next article.

Agentic misalignment
Hybris and artificial creation
Moral risks of AI

Continue reading in Global Order and Geopolitics