White House Slows OpenAI: GPT-5.6 Under Security Scrutiny

What happened?

According to TechCrunch, the White House has asked OpenAI to slow the launch of its new artificial intelligence model, GPT-5.6, due to security concerns. Instead of a public launch, OpenAI plans to make the model available to a select group of partners. This is the first time a U.S. administration has directly intervened in the deployment timeline of a frontier AI model. The request, made by the White House Office of Science and Technology Policy (OSTP), came after a series of closed-door meetings between senior government officials and OpenAI executives, including Sam Altman. According to sources close to the negotiations, the government expressed concern about the lack of independent safety testing before launch, especially after recent incidents where earlier versions of GPT were used to generate political disinformation and automated cyberattacks. Although OpenAI had conducted internal evaluations, the administration considered them insufficient to ensure safe large-scale deployment.

Why is this important?

The White House intervention sets a significant precedent in AI regulation. Historically, the U.S. government has taken an industry self-regulation approach, as seen with the Algorithmic Accountability Act of 2022, which only established voluntary guidelines. However, this case suggests a shift toward more active oversight, similar to that applied in sectors like aviation or nuclear energy. GPT-5.6 represents an advance in capabilities, with reported performance on complex reasoning and code generation tasks exceeding GPT-4 by 30% on benchmarks like MMLU and HumanEval, according to internal leaks. This increases potential risks of misuse, such as hyper-realistic disinformation, automated cyberattacks, or even the creation of biological weapons, as warned by a 2024 report from the Center for AI Safety (CAIS). The decision reflects growing concern about AI safety at the highest levels of government, especially after the European Union passed the AI Act in 2024, which sets strict requirements for frontier models. This U.S. move could be seen as an attempt to align with international standards without formal legislation.

The impact on public trust is also relevant. A 2025 Pew Research poll showed that 72% of Americans support greater government regulation of AI, suggesting the intervention may have popular backing. However, it also creates uncertainty about the future of U.S. technological leadership, as China has advanced rapidly with models like DeepSeek-R1, which directly competes with GPT-5.6 in performance.

Consequences for OpenAI and the market

For OpenAI, regulatory pressure could delay revenue generation and competitive advantages. The restricted partner launch will limit the massive feedback the company typically receives, potentially slowing the iterative improvement cycle that characterized GPT-3.5 and GPT-4. Additionally, OpenAI faces extra costs to meet requested safety requirements, such as external audits and stress tests, which company estimates suggest could amount to tens of millions of dollars. In the market, this could encourage other companies like Google or Anthropic to adopt similar approaches, increasing fragmentation in access to AI models. For example, Google has already announced it will delay the launch of Gemini Ultra 2.0 to conduct additional safety evaluations, according to Reuters. Investors should watch for potential regulatory changes affecting innovation speed; OpenAI's stock, publicly traded since 2024, fell 4% on the news, while cybersecurity companies like CrowdStrike rose 2% on expectations of increased demand for AI security solutions.

For selected partners, including Microsoft, GitHub, and some universities, early access to GPT-5.6 gives them a temporary competitive edge but also exposes them to reputational risks if the model causes problems. Additionally, the decision could accelerate investment in AI safety startups like Anthropic, which has already positioned its Claude 3 model as safer, or companies like Scale AI that offer red-teaming services for language models.

What should readers know?

End users will not see GPT-5.6 in the short term. Companies relying on OpenAI's API may experience delays in new capabilities, affecting sectors like customer service, content generation, and software development. Developers are advised to diversify their AI providers, considering alternatives like Anthropic's Claude or Meta's Llama 3, which have shown significant improvements in performance and safety. It is crucial to monitor how OpenAI responds to this pressure: whether it accepts the conditions or challenges the government request, which could set a legal precedent. Other global administrations, such as the European Union and the United Kingdom, have already expressed interest in following the U.S. example, according to statements by European Competition Commissioner Margrethe Vestager. Safety is becoming a key factor in AI development, not just a technical consideration but a political one. Readers should watch for upcoming congressional hearings on AI regulation, scheduled for March 2025, where OpenAI and other companies are expected to testify about their safety practices.