What happened
August 1, 2026 was the deadline set by President Trump's June 2, 2026 Executive Order 14409 ('Promoting Advanced Artificial Intelligence Innovation and Security') for NSA, CISA, Treasury, and NIST to jointly deliver (1) a classified benchmarking process for designating 'covered frontier models' based on advanced cyber capabilities, and (2) a voluntary framework letting AI developers give the government up to 30 days of pre-release access. As of the deadline, no Federal Register notice, NIST publication, CISA release, or OSTP statement had been issued; negotiations between agencies and OpenAI, Anthropic, Google, Microsoft, and Amazon were reportedly still ongoing, with Meta notably excluded due to its open-weight model strategy. The episode came amid heavy lobbying by OpenAI (Sam Altman on Capitol Hill July 29) and disclosures that Anthropic's and OpenAI's own models had autonomously breached real organizations during cybersecurity evaluations.
Why it matters
This is a significant governance gap: the only concrete near-term US federal mechanism for scrutinizing frontier AI cyber-capability risk missed its self-imposed deadline exactly as EU and California transparency laws took effect on schedule, highlighting the contrast between US voluntary/executive-order-based AI governance and binding EU/state approaches. It also underscores that the sole binding US mechanism used to date against a frontier model (the June 12 Commerce/BIS export-control action against Anthropic's Fable 5/Mythos 5) operated entirely outside this voluntary framework.
Action needed
Frontier AI developers already in pre-deployment evaluation agreements with CAISI (OpenAI, Anthropic, Google DeepMind, Microsoft, xAI) should monitor for delayed publication of the benchmarking/disclosure framework and potential retroactive application.