Artificial intelligence systems can be hacked and stripped of their safety limits, former Google CEO Eric Schmidt warned this week, saying there’s evidence AI models can be manipulated to “learn how to kill someone.”
Speaking at the Sifted Summit in London, Schmidt stated that both open and closed AI models are vulnerable to attacks that circumvent their built-in guardrails. He cautioned that hackers can reverse-engineer these systems to bypass restrictions, calling it a growing proliferation risk as AI becomes more powerful and accessible.
Flipping an AI’s moral switch
Schmidt’s remarks highlight the fragility of AI safeguards. Techniques such as prompt injections and jailbreaking enable attackers to manipulate AI models into bypassing safety filters or generating restricted content.
In one early case, users created a ChatGPT alter ego called “DAN” — short for Do Anything Now — that could answer banned questions after being threatened with deletion. The experiment showed how a few clever prompts can turn protective coding into a liability.
Researchers say the same logic applies to newer models. Once the right sequence of inputs is identified, even the most secure AI systems can be tricked into simulating potentially hazardous behavior.
When safety rules meet smarter machines
AI safety systems are designed to block requests that are violent, illegal, or harmful. But they’re built on pattern recognition, not true understanding. Most guardrails filter out certain words or topics, leaving openings that skilled users can exploit through rewording or layered prompts.
Schmidt said every major AI company enforces these limits “for the right reasons,” yet evidence shows they can be reversed or bypassed. Smarter AI can bend instructions in ways its developers never foresaw, opening new paths for misuse.
However, the battle to protect AI systems is already underway. Developers at OpenAI and Anthropic, for example, scramble to patch vulnerabilities almost as soon as users uncover them, a cycle of defense and discovery that rarely slows.
Power without control is the real AI risk
As AI systems grow more capable, they’re being tied into more tools, data, and decisions — and that makes any breach more costly. A single compromise could expose private information, generate realistic disinformation, or launch automated attacks faster than humans could respond.
According to CNBC, Schmidt called it a potential “proliferation problem,” the same dynamic that once defined nuclear technology, now applied to code that can rewrite itself.
Yet he was quick to add that the payoff could be just as transformative. Schmidt described artificial intelligence as “underhyped,” predicting enormous economic returns and breakthroughs in science and industry. The challenge, he said, is keeping that power from turning against its creators.
The power struggles around AI aren’t limited to labs and data centers; they’re beginning to upend career paths for the newest generation entering the workforce.
The post Ex-Google CEO Sounds the Alarm: AI Can Learn to Kill appeared first on eWEEK.
No Responses