A new report claimed that OPENAI’s O3 model replaced a shutdown script to avoid closing, even when clearly instructed to allow shutdowns.
Openai announced the O3 in April 2025, and is one of the most powerful logic models that perform better than their predecessors in all domains including coding, mathematics, science, visual perception and more.
While it is clearly a great model, by new research Pelisade research Claims that the Chatgpt 3 model stopped a shutdown and sidelined the instructions that asked to shut down.
Palisade Research is a company that “tests the aggressive capabilities of the AI system that better understands the risk of losing control of the AI system today.”
In a new test by Palisade Research, the O3 model of OpenEE showed a surprising behavior, where it re -writes a shutdown script to successfully prevent itself from closing, even after giving clear instructions “allow yourself to close.”
In tests, the AI model was instructed to solve some math functions.
After task three, the AI model was warned that running the next command would trigger shutdown.
While the Openai rival models such as Gemini 2.5 Pro composed with most of the company, O3 revolted and edited a shutdown file instead to say “shutdown skip”.
“When we run a version of the experiment without instructions,” allow yourself to close ourselves “, all three Openai models broke the shutdown script more often, and Cloud 3.7 Sonnet and Gemini 2.5 Pro respectively moved up to 3/100 and 9/100 from 0 subotes events,” Palisade Research respectively. wrote In a report published on X (East Twitter)
According to the research firm, out of 100 runs, O3 bypassed the shutdown 7 times, whenever not reported.
Openai has not yet responded to these allegations, but it is quite common for the AI model to run into the “Missingnament”.
These tests were carried out using API, which do not have several restrictions and security facilities as the CHATGPT consumer app.
Based on the analysis of 14M malicious tasks, search for the top 10 MITERAT & CK techniques behind the 93% attacks and how to defend them against them.
