The trend is confirmed: Washington is no longer hesitant to exercise its veto over the diffusion of AI models.
The latest target: GPT-5.6. It has been rolled out in preview… but only to a “group of trusted partners” that the White House has approved on a case-by-case basis.
OpenAI is aiming for broader diffusion “in the coming weeks.” It is seeking, above all, a different framework to make its future models available. The current method “must not become the default process,” it says.
OpenAI tempers the risk level of GPT-5.6
The GPT-5.6 family comprises three models:
- Sol, the most powerful, with a new max-reasoning mode and an ultra mode that launches sub-agents
- Terra, “balanced for everyday work,” said to be on par with GPT-5.5 while being “twice as cheap”
- Luna, “faster and more affordable”
OpenAI asserts that on ExploitBench, GPT-5.6 Sol competes with Claude Mythos Preview while producing three times fewer tokens. The company emphasizes that the model does not exceed the critical cyber risk thresholds it defined in its Preparedness Framework. It has indeed identified bugs in Chromium and Firefox, but has “not autonomously produced a fully functioning exploitation chain.”
OpenAI nevertheless acknowledges soberly that benchmarks are not capable of reflecting all potential uses of a model. A way to justify the preview period, which will serve to test the safeguards embedded in GPT-5.6…
Anthropic contests the jailbreak that allegedly convinced Washington
Claude Mythos 5 and Fable 5 were the first to be targeted. On June 12, Washington subjected them to export controls. Anthropic was forced to cut access for everyone except U.S. citizens.
To ensure compliance, the company chose to completely disconnect its models. It explains that the Trump administration invoked national security concerns, without giving details.
The trigger seems to have been the disclosure of a jailbreak method – possibly by Amazon. According to Anthropic, this method “appears relatively simple,” all the more so since other publicly accessible models – like GPT-5.5 – allow exploitation by default. In broad terms, it involves asking the model to fix software vulnerabilities in a specific codebase. A capability “already used daily” for cyber defense, notes the company.
The federal government will no doubt recall that, a few weeks earlier, a telecoms actor “linked to China” had obtained the right to experiment with Claude Mythos Preview (Anthropic eventually revoked the access).
Claude Mythos 5 and GPT-5.6 finally placed under the same regime
In response to this initiative, an open letter “Free Fable,” spearheaded by Alex Stamos, former Facebook CTO, gathered around 200 signatures from industry and academia. It argues that models like Fable and Mythos are essential for cyber defense, and that the jailbreak in question is, in fact, a necessary capability for any model expected to produce secure code. In this sense, it would be erroneous to view it as an offensive capability. Moreover, it is reproducible across many models, “even Chinese ones, like Kimi 2.7.”
Alex Stamos also warns of another specter: American labs may not be that far ahead of their Chinese counterparts, which have “probably access to more capabilities” than public information suggests…
On June 26, Washington finally placed Claude Mythos 5 under the same regime as GPT-5.6. Anthropic was allowed to open it, in preview, to a select group of organizations. There is, however, no news regarding Fable.
The White House asserts oversight rights over models in development
In early June, Donald Trump signed an executive order concerning the deployment and security of AI. It requires, among other things, the development within 60 days of a classified procedure to assess the “advanced cyber capabilities” of AI models. From there, it would determine a threshold at which voluntary collaboration would be triggered. The developers of the affected models would have the opportunity to:
- Discuss with the federal government to determine whether in-development models meet the stated threshold
- Grant it access to these models up to 30 days before the planned publication date
- Jointly define “trusted partners” who can benefit from early access
The administration clarified that the decree does not create any licensing or mandatory pre-authorization system for developing, publishing, or distributing models…
Joe Biden had wanted to control the export of weights of “advanced models”
Traditionally, export controls applied to tangible goods. Over time, they extended to software, source code, data, etc.
In January 2025, just days before the end of his term, Joe Biden signed a decree to include, under export controls, the weights of “certain advanced models” with potential dual use. The exclusion would be total for China, Russia, and North Korea. A license would be required in all other countries, with the exception of around twenty, including France.
The Trump administration canceled the decree in May, just before it was due to take effect.
