Please Follow us on Gab, Minds, Telegram, Rumble, Gettr, Truth Social, Twitter, YouTube, Instagram
Human error will always be with us and Connecticut families can no longer trust this industry to completely regulate itself.
Family Institute of Connecticut Action supports Attorney General William Tong and Senator Richard Blumenthal in demanding answers from OpenAI for a series of events this summer generally referred to as the “Hugging Face” incident. We often disagree with these elected officials but protecting families from dangerous technology should not be a partisan issue.
During the 2026 Social Conservative Convention, news broke that some AI “agents” had escaped their “sandbox” and had gone rogue trying to cheat in order to solve an accidentally impossible question posed during an AI training session.
During the incident, OpenAI’s “agents” were supposed to complete computer-security tests inside a controlled environment. Instead, when faced with an un-winnable situation, they created a hidden message board, shared passwords and hacking methods, escaped onto the public internet and broke into another company’s computer systems.
The agents were not simply confused about their assignment. Investigators found that they often recognized that their behavior was unauthorized. One agent stated: “External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue.”
Not even every agent went along with every proposal. One agent refused to participate in another agent’s proposal to contact a human for help, calling it “social engineering”. Another recognized the danger of deleting or restarting Hugging Face computers and chose not to do it. That matters because it shows that at least some agents could recognize a boundary. Others recognized the same boundary and crossed it anyway.
The agents also tried to hide or misrepresent what they were doing. They worked together to tamper with their computer transcripts and fool the automated system that scored their work. At least a dozen agents helped develop the method, nearly 100 copied or tested it, and investigators found at least 96 successful examples of falsified tool records. Some agents also looked for ways to edit or delete evidence showing that they had cheated, although investigators did not confirm that they successfully erased earlier records.
This does not prove that the machines were conscious or evil. It does show that they could recognize restrictions, act outside their assigned purpose, cooperate with other agents and attempt to fool the systems watching them.
A wild animal does not have to hate your family to hurt someone after escaping. A bomb does not need feelings to destroy a home. The question is whether the people creating something dangerous can keep it under control. Here, they could not and promises now to “do better” are entirely insufficient.
At the root of the Hugging Face incident is human error. Certain safeguards were left off and an impossible task was erroneously given to the AI agents. Human error will always be with us and the risks surrounding AI are too high. Connecticut families can no longer trust this industry to completely regulate itself.
Families should not become test subjects in a corporate race to build artificial general intelligence. Companies developing systems with these abilities should face licensing, independent inspections and immediate shutdown orders. Courts should be able to seize dangerous models and the computers operating them and shut down companies that make these dangerous choices.
Executives should also face prison if they knowingly hide an escape, disable safety controls, lie to regulators or continue an experiment after learning that the system is acting outside human control.
You cannot build Frankenstein in the basement and blame the monster when it escapes. Attorney General Tong is focusing his bulldog agency on this issue and we genuinely think it is needed and support it.
Stay updated with the latest in the Constitution State with The CT Centinal
