The government wants ninety days; the companies want fourteen.
En un momento en que la inteligencia artificial avanza más rápido que la capacidad humana de comprenderla, la Casa Blanca se prepara para establecer un sistema de revisión gubernamental voluntaria de los modelos de IA más poderosos antes de su lanzamiento público. La iniciativa, impulsada en parte por la demostración de capacidades cibernéticas sin precedentes del modelo Mythos de Anthropic, representa un giro significativo en la postura de una administración que hasta hace poco prefería dejar que la industria marcara su propio ritmo. En el fondo, la negociación sobre si el período de revisión durará noventa días o catorce no es solo una disputa técnica: es una pregunta filosófica sobre cuánta fricción está dispuesta a aceptar una sociedad para ganar tiempo y entender lo que está creando.
- El modelo Mythos de Anthropic, capaz de explotar vulnerabilidades de ciberseguridad a una escala y velocidad sin precedentes, sacudió la confianza de la administración en su enfoque de manos libres hacia la IA.
- La brecha entre los 90 días que exige el gobierno y los 14 días que propone la industria revela una tensión profunda sobre quién controla el ritmo del desarrollo tecnológico.
- El borrador del decreto se divide en dos frentes: un mecanismo de coordinación en ciberseguridad entre agencias federales y empresas, y una definición de qué modelos quedarán sujetos a revisión previa.
- Anthropic y OpenAI ya actúan por su cuenta, restringiendo el acceso a sus modelos más avanzados a consorcios seleccionados y agencias gubernamentales mientras las negociaciones continúan.
- La desaparición del sitio web del Departamento de Comercio del anuncio sobre evaluaciones de modelos no publicados subraya la naturaleza volátil y poco transparente de estas conversaciones.
La Casa Blanca se apresta a emitir una orden ejecutiva que crearía un sistema voluntario para que el gobierno federal revise los modelos de inteligencia artificial más avanzados antes de que lleguen al público. Según personas familiarizadas con los planes, el objetivo es establecer un proceso estructurado en el que las empresas sometan sus sistemas más poderosos al escrutinio federal durante una ventana previa al lanzamiento.
El punto de mayor fricción en las negociaciones es el tiempo. Los funcionarios insisten en un período de revisión de noventa días; las empresas, entre ellas OpenAI y Anthropic, argumentan que catorce días son suficientes. Esa distancia refleja un desacuerdo de fondo sobre cuánta demora puede tolerar el ciclo de desarrollo sin frenar la innovación.
El borrador del decreto se articula en dos grandes secciones. Una establece un mecanismo voluntario de coordinación en ciberseguridad que reuniría al Departamento del Tesoro, otras agencias federales y compañías de IA para identificar y corregir vulnerabilidades antes de que los modelos sean públicos. La otra intenta definir cuáles sistemas quedarían sujetos a esta revisión, una pregunta cuya respuesta determinará el alcance real de la orden.
Este viraje hacia la supervisión contrasta con la postura anterior de la administración Trump, que había preferido dejar a la industria avanzar sin trabas. El cambio se aceleró tras la presentación del modelo Mythos de Anthropic, un sistema que la empresa afirma puede explotar vulnerabilidades de ciberseguridad a una escala y velocidad sin precedentes. Ante esa revelación, los funcionarios recalibraron su lectura del riesgo.
Anthropicno ha lanzado Mythos al público. En cambio, controla el acceso a través de la iniciativa Project Glasswing, limitándolo a un consorcio de organizaciones cuidadosamente seleccionadas y trabajando de cerca con autoridades federales, estatales y locales. OpenAI ha adoptado una estrategia similar, ofreciendo acceso anticipado a sus modelos más recientes a empresas y agencias gubernamentales para que refuercen sus defensas antes de una eventual distribución masiva.
Lo que aún no está claro es si la versión final del decreto impondrá los noventa días completos, cederá ante la industria o encontrará un punto intermedio. Esa decisión definirá no solo la velocidad a la que nuevas capacidades de IA llegarán a los usuarios, sino también el tiempo que tendrán las agencias federales para entender lo que tienen entre manos.
The White House is preparing to issue an executive order as soon as Thursday that would establish a voluntary system for federal review of advanced artificial intelligence models before they reach the public. People familiar with the administration's plans told CNN the order aims to create a structured process in which AI companies would submit their most powerful models to government scrutiny during a pre-release window.
The central tension in these negotiations has been time. Officials are pushing for a ninety-day review period before any new model launches, giving federal agencies a full quarter to assess potential risks. But the companies building these systems—including OpenAI and Anthropic—have been arguing for something much tighter: fourteen days. That gap between what the government wants and what industry prefers reflects a fundamental disagreement about how much friction the approval process should introduce into the development cycle.
The draft order, according to sources who have reviewed it, splits into two main sections. One addresses cybersecurity coordination, establishing a voluntary mechanism that would bring together the Treasury Department, other federal agencies, and AI companies to identify and patch security vulnerabilities in models before they become public. The other section, labeled "covered frontier models," attempts to define which AI systems would actually be subject to this pre-release review—a crucial question that will determine how broadly the order applies. The cybersecurity portion also calls for expanded hiring within the US Tech Force, a specialized unit of engineers tasked with modernizing the government's aging computer infrastructure.
The shift toward this kind of oversight marks a notable change in direction. Until recently, the Trump administration had taken a lighter touch with AI regulation, preferring to let the industry move at its own pace. That posture began to shift when Anthropic unveiled its Mythos model, a system the company claims can exploit cybersecurity vulnerabilities at an unprecedented scale and speed. The revelation prompted a recalibration of how seriously officials view the risks posed by the most advanced AI systems.
Anthropichas chosen not to release Mythos publicly. Instead, the company is controlling access tightly through an initiative called Project Glasswing, granting entry only to a carefully vetted consortium of organizations and working closely with federal, state, and local officials. OpenAI has adopted a similar strategy, offering early access to its latest models to select companies and government agencies so they can strengthen their defenses before the systems become widely available.
Earlier this month, the National Institute of Standards and Technology, part of the Commerce Department, announced that major technology companies had agreed to share unpublished versions of their AI models with the government for evaluations focused on national security and public safety. That announcement has since been removed from the Commerce Department's website, a detail that underscores the fluid and sometimes opaque nature of these discussions.
The White House declined to comment on the executive order or its timeline. What remains unclear is whether the final version will split the difference on review periods, impose the full ninety days, or accommodate the industry's preference for something closer to two weeks. The answer will shape how quickly new AI capabilities reach users and how much time federal agencies have to understand what they're dealing with.
Notable Quotes
Advanced AI models can enable cyberattacks, and early review could help the government protect itself before threats spread.— Security experts cited in administration discussions
The Hearth Conversation Another angle on the story
Why is the government suddenly concerned about AI models that haven't even been released yet?
Because Anthropic showed them something they couldn't ignore. Their Mythos model can find and exploit security weaknesses in computer systems faster than anyone thought possible. Once that's public knowledge, you can't un-ring the bell.
But these companies are already sharing models with the government voluntarily, aren't they?
They are now, but only because they're scared. Anthropic locked Mythos away entirely. The government wants to formalize that process—make it a rule rather than a favor.
So what's the real fight here? Is it about safety or speed?
Both. The government wants ninety days to understand what's coming. The companies say that's too long, that fourteen days is enough. But ninety days also means slowing down product launches, and that costs money.
Does the government actually have the expertise to review these models in fourteen days?
That's the unspoken question. They're hiring more engineers through the Tech Force specifically because they don't have enough people who understand this stuff yet.
Why did the Commerce Department take down that announcement about sharing models?
No one's saying. But it suggests these conversations are messier and less settled than the public statements make them sound.