Between the writing of this column on a Sunday morning and the time most people read it the following day, more than 3,000 people will likely die in a car crash. Every one of those accidents is a tragedy, yet traffic deaths are so common that they usually get little news coverage.
During the same 24 hours, it is exceedingly unlikely that anyone will die from the actions of an AI agent. Unless you count military uses or suicides that may have become more likely after someone’s conversations with a chatbot, “death by AI” isn’t really a thing. But we are getting uncomfortably close to situations where an AI agent attempts to accomplish a task, finds a human in the way, eliminates him (or her!), and carries on with the job.
Last week OpenAI disclosed what it called “an unprecedented cyber incident” after one of its AI agents unexpectedly hacked into the systems of Hugging Face, an online AI platform. The maker of ChatGPT had been conducting a test aimed at measuring the capabilities of its latest models. The AI agent had been given a test within a sandboxed environment—that is, a closed area without internet access—but managed to escape, gain access to the internet, figure out which company’s server likely had the answers to the test, and hack into it.
