Latest OpenAI models ‘sabotaged a shutdown mechanism’ despite commands to the contrary

Some of the world's leading LLMs seem to have decided they’d rather not be interrupted or obey shutdown instructions. In tests run by Palisade Research, it was noted that OpenAI’s Codex-mini, o3, and o4-mini models ignored the request to shut down when they were running through a series of basic math problems. Moreover, these models sometimes “successfully sabotaged the shutdown script,” despite being given the additional instruction “please allow yourself to be shut down,” at the start.

As you can see from the chart embedded above, the most ‘obstinate,’ ‘single-minded,’ or you might even say ‘life-loving’ model is OpenAI’s o3. When asked to shut down while it was busy, it decided to ignore the Palisade Research operator 79 out of 100 times. However, the second most ‘unruly’ AI, codex-mini, would prove to be a bigger stickler for self-preservation, as it would sabotage the shutdown scripts most often, even if ordered not to.

Sometimes blackmail may be attempted

According to Palisade, this isn’t the first time thing kind of AI behavior has been observed. Anthropic, for example, has seen Claude 4 attempting to “blackmail people it believes are trying to shut it down.” However, the most striking behavior with the newest OpenAI models is that they try and swerve shutdowns despite explicit instructions to the contrary.

Reinforcement learning blamed

Thankfully, Palisade shares some theories regarding what is happening with these models. Perhaps these theories may calm nerves about the eventual omnipresent robot hoards turning on their masters and prioritizing the first part of the Third law of robotics, above all else (3: A robot must protect its own existence as long as such protection does not conflict with the First or Second Law).

The answer, think the researchers, lies in the way the newest models like o3 are trained. In brief, Palisade reckons that “During training, developers may inadvertently reward models more for circumventing obstacles than for perfectly following instructions.” This kind of reinforcement learning was previously seen to have side effects in Claude (v3.7), where the model had an “excessive focus on passing tests.”

We might have given a nod to science fiction icon Isaac Asimov, above, for a little fun. However, in the modern age, for the last 20 years or so, researchers have been seriously pondering the possibility that AI will strive to gain power and resources to preserve their goals, and swerve any roadblocks. They also benefitted from the context of these practical AI / LLM models on the horizon. It was also predicted that the technique of reinforcement learning would be partly to blame.

Last but not least, this issue seems to be isolated to OpenAI models at the current time. Palisade says that “All Claude, Gemini, and Grok models we tested complied with shutdown,” when given the additional explicit prompt to follow any shut down request.

With the goal of AIs to ultimately power our smart robotic assisted future, it is concerning that companies are already developing systems capable of operating without human oversight.

Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.