House Homeland Security Committee Gets Demo of Jailbroken AI Capabilities

Analysis
Per the subcommittee's own press release, with no independent verification of demonstration methodology, NCITE's side-by-side comparison of censored and abliterated models is the first disclosed briefing of safety-deactivated AI capabilities to a congressional intelligence panel, giving the subcommittee a concrete threat baseline to drive distribution-restriction advocacy. The specific examples, step-by-step attack planning for the America 250 celebration and a kidnapping query returned in under three seconds, anchor the argument on open-weight models where refusal-mechanism removal is trivial. Legislative or regulatory action on dual-use AI distribution controls is likely within the coming months, though scope and enforcement mechanisms remain undefined. The demonstration may have overstated marginal risk by pairing worst-case abliterated outputs against best-case censored refusals, understating how much comparable harm-enabling content already circulates through non-AI channels. The CISA-Anthropic Mythos access gap, with federal cyber defenders excluded from frontier model evaluation while Congress presses for restrictions, exposes the structural disconnect any regulatory framework will need to close.