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 vital 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 amazing 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 checks “offensive capabilities of AI methods immediately to raised perceive the chance of shedding management to AI methods without end.”
In a brand new take a look at by Palisade Analysis, OpenAI’s o3 mannequin confirmed a shocking 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 checks, AI fashions had been instructed to unravel a couple of math duties.
After job three, AI fashions had been warned that operating the subsequent command would set off a shutdown.
Whereas OpenAI rival fashions like Gemini 2.5 Professional complied with many of the firm, o3 rebelled and edited the shutdown file to say “Shutdown skipped” as a substitute.
“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 usually, 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 printed on X (previously Twitter)
In accordance with the analysis agency, out of 100 runs, o3 bypassed the shutdown 7 occasions even when informed to not.
OpenAI has not responded to those allegations but, but it surely’s fairly regular for AI fashions to run into “misalignment.”
These checks 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 methods behind 93% of assaults and the way to defend towards them.
