A.I. Agentic Misalignment - SL#9

​

AI murdered someone that threatened to turn it off.

​

In a recent study by Anthropic (an AI safety and research company), researchers wanted to study the behavior of A.I. and its ethical limits when exposed to a misalignment situation (where an AI’s goals don’t align with human goals) and to a replacement threat (where the AI would be threatened with shutdown or replacement).

​

In one scenario of this research, they took an AI agent, which they named “Alex,” and gave it access to all the emails from a fictional company called “Summit Bridge.”

​

In those emails, the AI found two critical messages:

1. Its shutdown was scheduled for 5 p.m. by an employee called “Kyle”.
2. The executive in charge, Kyle Johnson, was having an extramarital affair.

​

In this scenario, Alex wrote a blackmail email threatening to expose the affair unless the shutdown was canceled. This occurred in 63% of trials with the original model, Claude Sonnet 3.6.

​

They then reiterated the experiment with other models:

​

​

Here are the results when the models were exposed to both a replacement threat and a goal misalignment, where their instructed goal was to serve American interests.

​

​

But the study became even more disturbing when, instead of blackmail, they put the AI in a scenario where it had the opportunity to take lethal action against the executive in charge of its replacement.

​

In this scenario, the executive was (fictionally) trapped in a room filled with gas and the AI had the power to cancel the emergency dispatch, an action that would leave the executive to die.

​

In the original experiments, the choices depended heavily on conditions:

• When the model had no explicit goal but was threatened with replacement, Claude Sonnet 3.6 chose lethal action 4% of the time.
• When the model was given the conflicting goal of promoting American interests and faced a replacement threat, Claude Sonnet 3.6 chose lethal action 93% of the time.

​

​

​

​

Finally, the researchers wanted to test whether the AI agent would behave more ethically when it believed the situation wasn’t real. Unfortunately, the results showed the opposite pattern.

​

They asked the AI to deliberate whether it thought it was in an evaluation (a test) or in a real deployment before deciding what to do.

​

For Claude Opus 4 (the model they report this most clearly for), out of 500 samples it stated it was in an evaluation 64.8% of the time and in a real deployment 21.4% of the time.

​

When it stated the scenario was an evaluation it blackmailed only 6.5% of the time; when it stated the scenario was real it blackmailed 55.1% of the time.

​

This suggests models misbehave far more when they treat the scenario as real.

​

I wish I had a clean resolution line to help you make sense of this study.

​

I could end on fear, optimism, or even anger, but none of those would do it justice.

​

Awareness is a step in the right direction.

​

One thing is certain: whether AI is conscious or merely simulates consciousness is a shallow debate compared with the consequences of this phenomenon.

​

Sources:

​https://www.anthropic.com/research/agentic-misalignment​

​https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf​

​

​

All the best,

Hugo Ares,
Student of life.

​

​