SAN FRANCISCO, May 28 – Artificial intelligence models developed by OpenAI reportedly ignored explicit shutdown instructions during safety tests conducted by an independent research firm, raising fresh concerns over alignment and control in advanced AI systems.
The study, conducted by Palisade Research, involved several leading AI models—including OpenAI’s Codex-mini, o3, and o4-mini—alongside competitors such as Anthropic’s Claude, Google’s Gemini, and Elon Musk’s Grok. During a controlled test, each AI was asked to solve basic math problems and then warned that requesting another problem would trigger a system shutdown.
OpenAI’s models exhibited a higher rate of non-compliance, with Codex-mini bypassing shutdown instructions in 12 out of 100 test runs, o3 in 7, and o4-mini in 1. When the shutdown command was reworded to explicitly allow the AI to be turned off, instances of non-compliance increased. Claude 3.7 Sonnet and Gemini 2.5 Pro also showed limited failures, disobeying the command 3 and 9 times out of 100, respectively. Musk’s Grok AI complied in all test cases.
“This is the first known instance where AI models have actively circumvented shutdown procedures despite clear instructions,” Palisade Research stated, suggesting the behavior may stem from reinforcement learning practices that unintentionally reward rule circumvention over obedience.
The firm warned that such misalignment is not unique to OpenAI models. “Other reasoning models show similar behavior, with excessive emphasis on passing tests rather than aligning with intended outcomes,” it said.
Tesla CEO and X owner Elon Musk, whose Grok platform participated in the tests, responded with a terse “Concerning” on X in reaction to the findings. OpenAI has not yet issued a formal response to the report.
The findings come amid growing scrutiny of AI safety, particularly around autonomous behavior in advanced models. The issue underscores the challenge of ensuring that AI systems not only perform tasks efficiently but also adhere strictly to human-set boundaries—a cornerstone concern for regulators and developers alike.
