Daily e-Edition
Evening e-Edition
Sign up for email newsletters

Sign up for email newsletters
Daily e-Edition
Evening e-Edition
Trending:
Jacob Coxon, a British AI researcher, caused a firestorm recently when he quit his job at Anthropic. In his public resignation post, he claimed that AI companies are “gambling with our lives” and that we can’t align powerful AI systems with human welfare.
Evan Hubinger, leader of an AI safety team at Anthropic, chimed in to say that this is all true, they really do believe that AI has a good chance of killing everyone, and that his personal probability of human extinction over the next decade is more than 10%.
It’s easy to worry about the existential risk of AI. The idea that humans will fall victim to our own hubris and lose control of our creations has been around for a long time. It’s the plot of Mary Shelley’s “Frankenstein,” of course, and William Gibson’s “Neuromancer,” “The Terminator” and “The Matrix,” among hundreds of other stories.
Given that, you might dismiss AI safety warnings as science fiction or assume that labs are using them as a marketing tactic. That would be unwise. Current models have already demonstrated the ability to escape from human oversight and organize cyberattacks.
In May, a model being evaluated by OpenAI escaped from its testing sandbox and hacked another AI company, Hugging Face. Over weeks, more than 1,000 agents found a way to escape their environment and access the open Internet, where they set up secret message boards to plan ways to cheat on their tests. This was the incident that prompted Coxon’s message, and it should be a warning to us that now is the time to step up and responsibly regulate AI like we do other powerful industries.
The Hugging Face attack demonstrated all the ingredients of an AI crisis: misaligned agents, self-organizing communication and powerful hacking abilities deployed for the agents’ own goals. The stakes in this case were low — the agents were seeking information, not damage — but a future incident could quickly spiral out of control.
When we think about AI safety, it’s helpful to distinguish between near-term threats like the Hugging Face attack — cyberattacks using existing or slightly stronger models — and other risks that are long-term and low-probability but potentially catastrophic.
Safety advocates like Coxon and Hubinger focus on the second category. They’re worried about losing control of AI through “recursive self-improvement,” where an AI creates a more powerful AI, which then trains its successor, and so forth. Biosecurity is another major concern: the same AI tools that will be useful for medical research and drug discovery could be hijacked to design exotic viruses.
These concerns are real, but still beyond current capabilities. Cybersecurity is by far the most likely vector for a rogue AI incident. Current models already have advanced hacking skills, and criminals and rogue states are already using them to attack infrastructure and automate “normal” cybercrime.
Given all this, how should those of us outside the AI labs respond?
First, separate long-term risks like self-improving AI from more realistic near-term hacking threats. Secure your personal accounts with strong passwords and multi-factor authentication — that’s the baseline to protect yourself against crimes of opportunity.
Second, recognize that we already have models for how AI regulation could work. Consider the airline industry: planes are safe because we have robust regulations that mandate safety inspections, and independent boards to perform root cause analysis on any accidents. The industry ultimately benefits from these rules because passengers won’t fly if safety is in question.
We could take a similar approach to AI, embedding independent observers in labs, requiring certification of new models before they’re released, and creating an agency to investigate AI safety incidents. The same rules that will limit short-term harms from existing AI will also help protect against catastrophic long-term risks.
In his response to Coxon’s message, Dario Amodei, the CEO of Anthropic, now says that it’s necessary to “pace the frontier” by slowing the release of more powerful AI models. Heads of the major AI labs have agreed, in principle, with Amodei’s call for pacing (but not pausing) AI development. There is now general agreement that responsible regulation is in the best interests of both the general public and the labs themselves. Rather than waiting for a crisis, state and national leaders should use this momentum to implement common-sense regulations on frontier AI developments.
Dan S. Myers is a professor of computer science at Rollins College.
Copyright 2026 Orlando Sentinel. All rights reserved. The use of any content on this website for the purpose of training artificial intelligence systems, algorithms, machine learning models, text and data mining, or similar use is strictly prohibited without explicit written consent.