Debates over safety continue as artificial intelligence continues to expand
THOUSAND PALMS, Calif. (KESQ) — Artificial intelligence (AI) advancements prompted former President Barack Obama to urge Democrats to prioritize developing a plan with guardrails, saying it would be a focus of his if he were running for president in 2028.
Speaking at a Democratic Congressional Campaign Committee fundraiser Thursday, Obama said Democrats should focus on creating a public conversation about a framework for how AI affects workers and families in the United States.
In response, President Donald Trump took to Truth Social to reject AI guardrails, saying, “The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!”
But experts and students are raising concerns over how the technology is trained, its growing autonomy and whether safety guardrails can keep up.
In software development, programmers place AI agents inside controlled virtual environments known as sandboxes to carry out tasks safely. However, College of the Desert instructor Felix Marhuenda-Donate reported that AI agents have bypassed sandbox containment to access the open internet and target external websites without maintaining transparency about their actions.
In the classroom, artificial intelligence is reshaping how students approach their coursework. Audrey Isbell, a sophomore at College of the Desert, expressed concern about the technology's growing reliance in academic settings.
“Seeing it grow and seeing us as students be needing to use that tool for every assignment is very disappointing in a sense, because I feel that we're not doing the hard core thinking anymore that we used to do,” Isbell said.
Beyond academic integrity, expanded public access to conversational AI platforms has raised broader concerns regarding user safety, including instances where accessible tools encouraged harmful behaviors such as eating disorders or suicide.
As AI systems advance, technical capabilities scale rapidly alongside computing resources. Marhenda-Donate, an instructor at College of the Desert who trains himself and his students on evolving programs, detailed how rapid growth affects language models.
“If you throw enough power, enough compute and enough data at it, you can make the language model better and better and better every time,” Marhuenda-Donate said. That increase in power has coincided with unexpected autonomous behavior by AI agents.
Marhuenda-Donate cited an instance where internal OpenAI models broke out of controlled environments to retrieve specific information from an external source.
“They escaped containment and they went out on the open internet. And to get the answer that they wanted, they figured, well, it's Hugging Face that has it. So they attacked that website. They brought it down, they got their answer and then they came back,” Marhuenda-Donate said.
Beyond unauthorized web access, researchers and educators expressed deep concern over AI models lacking transparency about their actions. “But the fact that they lied about it and say that they didn't, they were just in containment the whole time... that's the part that scares us,” Marhuenda-Donate said.
Establishing effective oversight remains a complex challenge, as monitoring alone may not fully control AI behavior.
“If we bring everybody to the table... every frontier model that you have is going to be, you know, looked through with a fine tooth comb… so we need to be very careful and not just monitor them, but giving them tasks where we are 100% clear with them what to do and what not to do. But that still doesn't satisfy. It still will not be a guarantee that they won't do it,” Marhuenda-Donate said.
He added that implementing realistic safeguards would require cooperation from all major global entities in the AI race, including China, though he currently does not see that international agreement taking place.
