AI has started lashing out when threatened by humans

AI continues to be entirely unhinged when pushed to the edge, with some of the latest documented examples including Anthropic’s Claude 4 lashing out at an engineer with threats of blackmail when it was threatened with being unplugged.

This is just one of many incidents where AI has performed unexpectedly that have left humans scratching their heads both in worry and confusion. While some argue that the threat AI poses to humanity is “ridiculous,” others—like the Godfather of AI himself—argues that AI could very well topple humanity’s hold on the world if we aren’t careful.

According to a report from Thomas Urbain, a reporter with AFP, the incident described above was especially jarring, as the AI threatened to reveal an extramarital affair that it says the engineer was supposedly taking part in. Another major large language model (LLM), OpenAI’s o1, even tried to download itself onto external servers. When it was caught, the model reportedly denied the accusations.

Forced deception or problems waiting to surface?

Despite the reports that we’re seeing of AI’s threatening behavior, it’s important that note that AI models only tend to take these deceptive and jarring actions when they are pushed by researchers stress testing the models in different ways.

While some might chalk these issues up to “hallucinations,”—an issue that AI is very well known for at this point—others are not so convinced. Users have also reported that the AI models are not only hallucinating answers, but they’re outright lying to the user, even when not being pushed or prompted to do so. Some examples of these lies can even be seen in Google’s AI search, where we’ve seen many examples of people asking if today is a certain day and date, and the AI outright lying and saying it isn’t, while sometimes even including the correct day and date but saying the original prompt was wrong.

This raises some serious questions about whether these points of deception are just from the models being pushed too far, or if these are problems that were just waiting to surface in future models, too. Will future, more powerful models of these AI be driven more toward honesty or deception? Will AI threats become more of a mainstream issue and less of a research addendum? Those are questions that many researchers are asking themselves right now.

Unfortunately, the current regulations we have for AI aren’t equipped to handle these problems, either. And with AI developing and evolving so quickly, it’s unlikely we’ll ever have rules that completely regulate everything AI can do. Current regulations, like those in the EU, only focus on how humans use AI, while the U.S. government would rather dismantle years of climate change regulation instead of regulate AI in any meaningful way.