A closed-door session for lawmakers showed how easily AI safety guardrails can be bypassed, raising urgent questions about regulation and national security.
The Department of Homeland Security showed members of Congress something they weren’t prepared for: AI models, once stripped of their safety guardrails, generating detailed plans for bomb-building and terror attacks with alarming speed.
The closed-door demonstration, organized by DHS’s National Counterterrorism Innovation, Technology, and Education Center (NCITE), took place over three days from April 22-24, 2026. It was designed to give lawmakers a visceral understanding of what “jailbroken” AI actually looks like in practice.
What lawmakers saw behind closed doors #
NCITE showcased both US and foreign AI models that had been modified to bypass built-in safety measures. The models demonstrated could rapidly generate harmful plans once their guardrails were disabled.
House Homeland Security Subcommittee Chairman Andy Ogles described what he saw as “frightening.”
From closed doors to public hearing #
The April demonstration set the stage for a public hearing on June 4, 2026, convened by the House Homeland Security Subcommittee on Cybersecurity and Infrastructure Protection. The hearing broadened the conversation beyond jailbroken chatbots to address frontier AI, agentic AI, and coding tools that could be weaponized by malicious actors.
Witnesses included representatives from Google Threat Intelligence and the Electronic Frontier Foundation.
Key figures from NCITE flagged a troubling trend during the proceedings: extremist groups are increasingly utilizing uncensored AI technology, making potential attacks more feasible and accessible than they’ve ever been.
The hearing also fits into a larger congressional investigation into AI models developed in China, adding a geopolitical dimension to what’s already a complicated domestic policy question.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our