The directive, reported by Bloomberg and Politico ahead of signing, would require federal review of certain frontier AI models before public release and impose new security requirements on defence-adjacent infrastructure. The rationale is straightforward: increasingly capable AI systems present novel attack vectors, particularly where they could accelerate offensive cyber operations or be targeted for theft by state actors.
Camellia Chan, CEO and co-founder of hardware security firm X-PHY, argues that the intent is sound but the implementation creates its own risks. Concentrating access to the most capable AI models inside government review environments makes those environments immediate targets for espionage. If software-layer controls remain the primary defence, a single successful compromise could mean full-system access.
Washington's push for greater oversight of frontier AI models reflects growing concern over these systems' cyber capabilities. Anthropic's decision not to publicly release Mythos last month has only reinforced fears about how quickly advanced AI could accelerate offensive cyber operations and attacks against critical infrastructure.
The concern extends to the broader trajectory of AI deployment in critical sectors. Moving frontier models into defence and infrastructure contexts before the underlying security architecture is ready inverts the sequence. Hardware Root of Trust — security that is enforced at the chip level rather than in software — becomes critical when the systems being protected are themselves capable of generating novel attack methods.
The AI cybersecurity order follows several months of discussion in Washington about how to handle frontier model development. Anthropic's decision not to release its Mythos model publicly was cited by multiple observers as evidence of how quickly the security calculus around advanced AI is shifting.
