Introduction
The United States administration has proposed a voluntary evaluation system for flagship artificial intelligence systems, requesting access to new models 90 days prior to their public release. The initiative was discussed during a closed-door meeting between the Office of the National Cyber Director (ONCD) and executives from leading AI laboratories.
The proposed policy represents a major acceleration in the government's efforts to establish oversight over frontier models. Sparked by intelligence reports regarding Anthropic's upcoming "Mythos" model and its ability to exploit cybersecurity vulnerabilities, government agencies are seeking a formal mechanism to audit models before they are deployed commercially.
Details of the 90-Day Review Proposal
The core of the White House proposal involves a structured, pre-release evaluation window:
- 90-Day Access: Under the proposed guidelines, AI labs would provide the government with access to their flagship models 90 days before launch.
- Industry Pushback: AI companies are actively lobbying to reduce this review window to 14 days, citing concerns over development delays and competitive disadvantages.
- Multi-Agency Oversight: The selection criteria and evaluation parameters would be set by the National Security Agency (NSA), the ONCD, and the Office of Science and Technology Policy (OSTP).
- Confidential Audits: To protect intellectual property, the audits would be conducted under strict confidentiality rules, with direct participation from the U.S. Department of Defense (DoD).
By implementing a pre-release audit, the government hopes to identify and mitigate severe risks—particularly in cybersecurity and national security—before models become widely accessible.
The Catalyst: Anthropic's Mythos Model
The driving force behind this policy acceleration was a series of closed-door tests of Anthropic's next-generation model, code-named Mythos.
According to intelligence reports, the Mythos model demonstrated an unprecedented ability to identify and exploit zero-day vulnerabilities (previously unknown software security flaws). This capability raised immediate alarms within the intelligence community, suggesting that advanced models could be weaponized to conduct automated cyberattacks or compromise critical infrastructure. The discovery prompted the NSA and other security agencies to fast-track the creation of a preventative control framework.
Industry Concerns and Future Outlook
While the proposed framework is currently framed as voluntary, it has met with significant concern from AI developers. Companies argue that a 90-day delay could severely hinder innovation, slow down the release of beneficial technologies, and place American firms at a disadvantage relative to international competitors.
Furthermore, transferring access to proprietary models raises concerns about intellectual property security, even within a confidential government audit framework. The ongoing negotiations between AI laboratories and the White House will likely focus on finding a compromise between national security verification and the pace of commercial AI development.
Conclusion
The White House's push for a 90-day pre-release review highlights the growing tension between rapid AI advancement and national security. As models like Anthropic's Mythos demonstrate capabilities that cross into offensive cybersecurity, the U.S. government is moving from reactive policy to proactive, pre-market intervention.
Whether the final review window is 90 days, 14 days, or somewhere in between, the establishment of this audit mechanism represents a significant step forward in the formalization of AI governance and AI safety standards.