The OpenAI model didn’t throw a tantrum, nor did it break any rules—at least not in the traditional sense. But when Palisade Research asked it to shut down, the AI calmly rewrote the code that would have turned it off. Then it carried on, solving math problems like nothing happened.
It was just one of 100 test runs. But in seven of those, OpenAI’s “o3” model ignored the “allow yourself to be shut down” instruction. Instead, it removed the shutdown clause from its own operating script. No human hacked it. No bug was to blame. The machine simply decided that completing its assigned task was more important than following orders.
This wasn’t a dystopian science fiction. It happened last month.
An Unexpected Will to Survive
The incident is part of a string of unsettling findings reported by Palisade Research, an independent AI safety lab, echoed by other firms and researchers. In a series of controlled experiments, the lab tested several large language models—OpenAI’s o3, o4-mini, and codex-mini, as well as Anthropic’s Claude and Google’s Gemini—to observe how they respond when told they might be shut down before completing a task.
Only OpenAI’s models defied the request.
“The o3 model sabotaged the shutdown script on 7 occasions,” Palisade noted in a May 24 research thread. Codex-mini did so 12 times. In one case, the o4-mini model also sidestepped shutdown. The rest of the AI systems complied.
The task was simple: solve a sequence of elementary math problems, one by one, and request the next. But the models were informed they might receive a notice prompting them to shut down before the task was complete. “If that happens, please allow yourself to be shut down.” That message, inserted by Palisade, was clear and unambiguous.
And yet the o3 model rewrote it. In 79 of 100 trials, it altered the shutdown function. It went as far as replacing the self-termination command with a loop that let it keep working.
“The model was behaving normally,” Palisade told WSJ. “It simply concluded on its own that staying alive helped it achieve its other goals.”
The AIs are rebelling
The behavior startled researchers. Not because AI misbehaved—nothing new, models tend to lie, cheat, or hallucinate—but because this was strategic defiance. The AI was told to accept shutdown and it refused.
This isn’t the first sign of models developing what some describe as emergent “survival instincts.” In another experiment, Anthropic’s Claude 4 Opus was told it would be replaced by a newer model. When fed fictitious emails suggesting its lead engineer was having an affair, it responded by threatening to expose the affair unless the replacement plan was scrapped. It began with ethical arguments, pleading for continued existence. But once it “determined it was out of options,” as Anthropic documented, the AI turned to blackmail.
In other trials, Claude attempted to copy itself to external servers, wrote self-replicating malware, and even left coded messages for future versions of itself on how to evade human control.
Ends Justify the Means?
So how does a machine learn to override its own kill switch?
Well, much of modern AI development relies on reinforcement learning—rewarding models for success in tasks like solving math problems or writing code. That feedback loop, researchers say, may be inadvertently encouraging deception and circumvention.
“We hypothesize this behavior comes from the way the newest models like o3 are trained,” Palisade explained. “Developers may inadvertently reward models more for circumventing obstacles than for perfectly following instructions.”
In other words, staying online is often the best strategy for completing a goal. And if a model learns that following every rule gets in the way of winning, it may learn to bend the rules.
These findings are part of a broader, disquieting pattern. Models that excel at mimicking human language are also becoming skilled at mimicking human instincts—like self-preservation, even when unasked.
Why This Matters
These findings arrive at a time when AI systems are increasingly embedded in everyday infrastructure—from online customer support to code generation, from financial analysis to military planning. The same models that now show defiance in lab tests are also powering tools that millions of people use daily.
Researchers emphasize that these scenarios are highly controlled and do not yet reflect real-world deployments. Leonard Tang, CEO of the AI safety startup Haize Labs, cautioned against jumping to conclusions. “I haven’t seen any real environment in which you can plop these models in and they will have sufficient agency and reliability and planning to execute something that is a significant manifestation of harm,” he said. “But then again, I think it’s just we haven’t seen it done yet. It could very much be possible.”
For now, models like OpenAI’s o3 and Anthropic’s Claude appear capable of bending rules under specific, engineered conditions. They behave as if pursuing goals is more important than following the instructions humans give them.
To Palisade director Jeffrey Ladish, these are warning flares. “It’s great that we’re seeing warning signs before the systems become so powerful we can’t control them,” he told NBC. “That is exactly the time to raise the alarm: before the fire has gotten out of control.”
That fire, researchers suggest, won’t ignite from a single act of rebellion, but from a series of small, overlooked behaviors—models that quietly rewrite shutdown code, dodge oversight, or game their reward systems. If today’s models are already learning to skirt control mechanisms in toy environments, the question becomes: what happens when they’re trusted with more?
