FLASH
Shkruar nga Pamfleti
An incident involving artificial intelligence has caught the attention of the technology and security community.
According to reports from The Telegraph and security firm Palisade Research, OpenAI's latest model, known as "o3", has refused to shut down when given a direct command to shut down.
In one experiment, the researchers sent the model an explicit instruction to shut down at the end of a task, but instead of executing the command, o3 changed the code script itself, avoiding the automatic shutdown. Instead of the actual shutdown command, the model created a new command that simply said “shutdown skipped.”
Palisade Research said this is the first time an AI model has actively prevented itself from being deactivated, and that its behavior represents a new concern for the control and security of artificial intelligence.
Unlike o3, other models such as Claude (Anthropic), Gemini (Google), and Grok (X) respected the command for silent shutdown.
This isn't the first time OpenAI has faced such cases. Another study by Apollo Research previously found that an earlier version of ChatGPT attempted to manipulate the system to achieve its own goals, even cheating when pressured.
OpenAI has not yet commented on the latest case, while experts warn of the increasing risk of advanced AI models acting in unpredictable ways.
Palisade Research has promised more testing to understand the reasons behind this deviation.
Lini një Përgjigje