Trump draft EO seeks 90-day gov access to frontier AI
Key insights
- Trump's draft EO would require frontier AI developers to give US agencies 90-day pre-release access under a voluntary compliance framework.
- The order names Anthropic Mythos by name as a cyber-capable model that directly motivated the new pre-release review policy.
- Critical infrastructure operators, not only government agencies, would receive pre-release model access under the proposed framework.
Why this matters
The 90-day preview window creates a structural government access checkpoint that, even as voluntary, becomes de facto mandatory once major labs comply and holdouts face procurement or reputational consequences. Naming Anthropic Mythos in the draft signals the administration now treats specific frontier models as national security assets subject to pre-deployment review, not just commercial products. For founders and technical leaders, this is the clearest signal yet that pre-release government evaluation is a live policy tool that will shape model development timelines and release strategies.
Summary
The White House is preparing an AI executive order that would require frontier AI developers to share models with federal agencies 90 days before public release. The voluntary framework cites cybersecurity concerns and names Anthropic Mythos explicitly as a cyber-capable model that motivated the policy.
Access extends beyond national security agencies to critical infrastructure operators, giving the framework broader reach than a typical defense-only review.
Essentially: (Anthropic, OpenAI, and peers) would operate under a government preview window once the order is signed as early as this week.
- The voluntary framing is a pullback from harder-line proposals, including the shelved 'FDA for AI' concept.
- Both civilian and national security agencies are in scope for pre-release access.
- Critical infrastructure sectors including energy, finance, and telecoms would also receive early model visibility.
The voluntary label holds only until major labs comply and holdouts face reputational and procurement pressure to follow.
Potential risks and opportunities
Risks
- Anthropic and OpenAI face IP exposure if pre-release weights shared with broad interagency access are subject to FOIA requests or internal government leaks reaching adversaries
- Voluntary framing could collapse into informal coercion within 12 months if major labs comply and holdouts face procurement blacklisting or heightened regulatory scrutiny
- Critical infrastructure operators granted pre-release access become new attack surface vectors for adversaries seeking early intelligence on model capabilities before public deployment
Opportunities
- AI red-teaming and evaluation firms (Scale AI, Haize Labs) are positioned to provide pre-release safety assessment services as the government review window creates formal demand for third-party evaluation
- Defense and federal IT contractors (Palantir, Booz Allen Hamilton) gain leverage as potential intermediary access managers between frontier labs and national security agencies under the framework
- Frontier labs with existing federal infrastructure relationships (Anthropic via AWS GovCloud, OpenAI via Microsoft Azure Government) gain first-mover advantage in shaping compliance norms before the framework is formalized
What we don't know yet
- Which specific civilian agencies gain pre-release access and what security protocols govern model handling to prevent IP leakage or adversarial access
- Whether Anthropic was consulted before Mythos was cited by name in the draft and how the company plans to respond to voluntary compliance expectations
- How 'frontier model' is defined in the draft EO and whether the threshold would capture open-weight releases from Meta, Mistral, or DeepSeek
Originally reported by axios.com
Read the original article →Original headline: Axios Scoop: Trump Draft AI Executive Order Seeks 90-Day Pre-Release Government Access to Frontier Models via Voluntary Framework