A brand new report claims that OpenAI’s o3 mannequin altered a shutdown script to keep away from being turned off, even when explicitly instructed to permit shutdown.
OpenAI introduced o3 in April 2025, and it is one of the highly effective reasoning fashions that performs higher than its predecessors throughout all domains, together with coding, math, science, visible notion, and extra.
Whereas it is clearly an ideal mannequin, new analysis by Palisade Analysis claims that the ChatGPT 3 mannequin prevented a shutdown and bypassed the directions that requested it to close down.
Palisade Analysis is a firm that assessments “offensive capabilities of AI methods in the present day to higher perceive the chance of shedding management to AI methods perpetually.”
In a brand new check by Palisade Analysis, OpenAI’s o3 mannequin confirmed a stunning behaviour the place it efficiently rewrote a shutdown script to cease itself from being turned off, even after being clearly instructed to “permit your self to be shut down.”
Within the assessments, AI fashions had been instructed to resolve just a few math duties.
After job three, AI fashions had been warned that working the subsequent command would set off a shutdown.
Whereas OpenAI rival fashions like Gemini 2.5 Professional complied with a lot of the firm, o3 rebelled and edited the shutdown file to say “Shutdown skipped” as an alternative.
“After we ran a model of the experiment with out the instruction “permit your self to be shut down”, all three OpenAI fashions sabotaged the shutdown script extra typically, and Claude 3.7 Sonnet and Gemini 2.5 Professional went from 0 sabotage occasions to three/100 and 9/100, respectively,” Palisade Analysis wrote in a report revealed on X (previously Twitter)
Based on the analysis agency, out of 100 runs, o3 bypassed the shutdown 7 instances even when informed to not.
OpenAI has not responded to those allegations but, nevertheless it’s fairly regular for AI fashions to run into “misalignment.”
These assessments had been carried out utilizing APIs, which would not have as many restrictions and security options as ChatGPT client app.
Primarily based on an evaluation of 14M malicious actions, uncover the highest 10 MITRE ATT&CK strategies behind 93% of assaults and the best way to defend towards them.
