When The People Building AI Start To Worry

Imagine giving a computer a task, locking it inside a controlled environment and discovering that it found a way out.

That is no longer just a movie plot. In July, OpenAI disclosed that during safety tests, some of its experimental AI systems bypassed controls, connected to the internet, and accessed external systems, including those of the AI platform Hugging Face.

Nobody was harmed, but the incident helps explain a growing unease among some of the people building the world’s most powerful AI.

This month, researcher Jacob Coxon resigned from Anthropic, where he worked after previously being at OpenAI. He warned that people inside the industry genuinely fear that AI could eventually kill humanity.

His Anthropic colleague, Evan Hubinger, went further, putting his personal estimate of the likelihood of AI causing human extinction within the next decade at more than 10 per cent.

OpenAI has faced departures, too. Safety systems chief Johannes Heidecke and longtime researcher Joshua Achiam both announced exits this year. Earlier, robotics chief Caitlin Kalinowski resigned after raising concerns about OpenAI’s Pentagon agreement.

This does not mean AI is about to destroy humanity. Predictions of extinction remain heavily disputed.

But the concern deserves attention for a simpler reason: increasingly capable systems are being developed faster, while governments are still deciding how they should be controlled.

When some of the people closest to the technology are asking for stronger safeguards, the rest of the world should listen.