President Trump is expected to sign an AI cybersecurity directive that expands existing federal cybersecurity information-sharing programs to include AI companies. The directive stops short of mandatory federal approval for frontier-model deployments, an option that had been under discussion since the Anthropic Mythos disclosure earlier this month. The signing is expected in the coming days.
What is proposed versus what is rejected matters for builders. The directive expands information-sharing under existing CISA and ODNI authorities, which means AI labs would participate in classified threat briefings and contribute incident data to federal databases. Participation appears voluntary in the current draft, with incentives such as faster security clearances for AI lab personnel and access to classified threat intel. Penalties for nonparticipation are not in the draft. The proposed mandatory pre-deployment approval for frontier models, which had been the most aggressive option under discussion after the Mythos disclosure, was dropped from the directive. The administration apparently concluded that mandatory approval would create operational bottlenecks and was unlikely to survive legal challenge from labs already operating commercial products. State-level activity continues independently. California, Colorado, Texas, and Illinois all have active AI laws on the books. The federal directive does not preempt them; the patchwork compliance burden for frontier labs continues to grow regardless of what the federal directive includes. For labs operating below the frontier threshold, the directive imposes no direct obligations. For labs at or above the frontier threshold (10^26 FLOPs training compute or similar), the practical effect is access to classified threat intel in exchange for voluntary incident reporting. The trade is favorable for the labs and operationally minimal for the federal government.
For frontier-lab compliance teams: the directive''s voluntary structure means the operational burden is lighter than the Mythos-era discussion suggested. Continue building documentation infrastructure for what eventually becomes mandatory. For builders below the frontier threshold: federal AI policy now affects you only indirectly; the state-law patchwork remains the immediate compliance question.