Claude Mythos too powerful to ignore: White House seeks deal with Anthropic

Can a technology deemed too risky become a strategic asset overnight? In Washington, the Anthropic case illustrates this shift. After being excluded from federal systems, the AI ​​enterprise is returning to the heart of discussions, while its capabilities trigger as much interest as concerns. Between military requirements, cybersecurity issues and ethical limits claimed by its leaders, the American administration is now seeking a path to compromise in an issue that has become highly sensitive.

A ban contested behind the scenes in Washington

In February, Donald Trump's administration classified Anthropic as “a risk to supply chain security”ordering federal agencies to immediately cease all use of its technologies, with a six-month deadline to dispose of them.

This decision follows CEO Dario Amodei's refusal to grant unrestricted access to models for military uses. The president has justified this position by stating: “The United States of America will never allow a radical left-wing, so-called woke corporation to dictate how our great military fights and wins its wars!” ».

Now the White House is working on a directive that could change that trajectory. According to several sources, the goal is to allow federal agencies to reuse Anthropic's solutions while avoiding political disavowal. This development takes place in a context of intensive discussions between political leaders and economic players, a sign that the issues now go beyond the simple technological framework.

Here is some factual elements :

• The administration had imposed an immediate ban with a six-month deadline to abandon Anthropic;
• The conflict finds its origins in the refusal to open the models to unrestricted military uses;
• New directive would aim to circumvent 'supply chain security risk' classification;
• Meetings involved key figures like Susie Wiles and Scott Bessent as well as major banks.

Mythos, an AI at the heart of security tensions

At the center of this potential turnaround is Mythos, Anthropic's most advanced model unveiled in March. Designed to detect and exploit software vulnerabilities, it is already attracting the attention of public and private institutions.

The AI ​​Security Institute reports that Mythos Preview is the first model capable of succeeding thanks to on-chain data “The Last Ones” (TLO), a 32-step network attack simulation that typically requires 20 hours of human work. For its part, Mozilla asserts having identified and corrected 271 flaws in its browser using this technology, evoking an effect of astonishment in the face of the scale of the discoveries: “many other teams now feel the same giddiness as us when these results started to become clear”.

These capabilities explain the growing interest of government agencies, despite official restrictions. The National Security Agency is already using Claude Mythos Preview on classified networks, a sign of pragmatic adoption in the face of the challenges. At the same time, large financial institutions such as Citigroup or Goldman Sachs have been involved in discussions on the risks represented by such AI, capable of both securing and compromising critical systems.

The evolution of this issue could redefine the relationship between technology companies and states, as the capabilities of artificial intelligence exceed traditional frameworks of control. Between national security requirements and ethical red lines set by developers, the compromise remains fragile. Washington's next decisions will be closely scrutinized, as they could influence the global governance of AI and its uses in sensitive areas.