Senators Jim Banks, a Republican from Indiana, and Kirsten Gillibrand, a Democrat from New York, have introduced the Insider Threat Reporting and Security Guidance Act of 2026, an eighteen-page bill that would turn the Pentagon's relationship with its frontier AI suppliers from a procurement transaction into a standing oversight regime. The core mechanism is a reporting obligation on what the bill calls covered contractors: any company holding Department of Defense AI agreements worth one hundred million dollars or more. Within one hundred and eighty days of enactment, the Secretary of Defense would have to issue regulations defining what those contractors must disclose.
The disclosure list is unusually specific for AI legislation. Covered vendors would report their security policies, who has access to model weights and to training pipelines, and any unauthorized access, data exfiltration or sabotage. More novel, they would also have to disclose model behavior: past instances of safeguard evasion, unprompted autonomous actions, and other concerning conduct observed in their systems. Everything submitted would have to be recertified as accurate and complete at least every ninety days. Incidents carry hard clocks. A national security incident, with theft of model weights named as the canonical example, must be reported within seventy-two hours of discovery. Material vulnerabilities or concerning model conduct must be reported within seven days of the contractor determining that they are material, and the Secretary must brief Congress within seven days of an incident.
The bill builds on Gillibrand's Secure and Accountable Military AI Act from June, which set up the notification framework, and arrives as the department's dependence on commercial frontier models has deepened quickly. In 2025 the Pentagon awarded four frontier labs contracts worth up to two hundred million dollars each, and in May eight companies, Anthropic not among them, agreed to deploy models on classified networks. Gillibrand framed the bill as putting commonsense guardrails around that deployment.
Why it matters technically: the bill treats model behavior, not just information security, as a reportable supply-chain risk. A requirement to disclose safeguard evasions and unprompted autonomous actions effectively asks vendors to share the output of their internal alignment and red-team evaluations with the customer on a recurring schedule, which no current federal contract does in a standardized way. The one-hundred-million-dollar threshold confines it to the largest awards, so smaller integrators and fine-tuners would sit outside it. The caveats are the obvious ones: this is a proposed bill with no committee action yet, the regulations it mandates would take six months to write after passage, and terms like concerning behavior would be defined in those regulations rather than in statute.