The U.S. government is restricting access to high-power AI models from Anthropic and OpenAI to mitigate national security risks, according to reporting by NPR. These measures include targeted export bans on Anthropic’s Mythos-based models and a new agreement allowing the administration to screen users of OpenAI’s latest technology.
Why is the U.S. government restricting AI access?
Federal officials are treating cutting-edge AI models as potential national security risks because of their ability to identify and exploit software vulnerabilities. While these tools help defenders secure networks, they can also be used by bad actors to hijack data or crash infrastructure. This shift toward tighter control follows an increase in AI-assisted cyberattacks using less powerful, widely available tools, according to NPR tech correspondent John Ruwitch.
The scale of this capability is significant. Anthony Grieco, head of security at Cisco, told NPR that Anthropic’s Mythos model scanned 1.8 billion lines of code across multiple languages in eight weeks. Grieco noted that completing this task with previous human-machine combinations would have taken more than eight years.
What happened with Anthropic’s Mythos models?
In early June, Anthropic released two AI models based on the Mythos architecture. Although these models included built-in guardrails to prevent malicious cyber activity, the Trump administration informed Anthropic that those safeguards could be bypassed. Consequently, the government imposed an export ban, prohibiting foreigners—including some Anthropic employees—from using the models.
The Commerce Department partially lifted this ban on Friday. According to a letter from the commerce secretary viewed by NPR, the government now permits a specific list of American companies and their foreign staff to access the model. Anthropic stated in an email to NPR that it is pleased with the change and will continue collaborating with the government.
How is OpenAI responding to government oversight?
OpenAI has agreed to let the U.S. administration vet a list of companies granted access to its newest model. The company stated this screening process represents the most effective path toward making the model broadly available in the coming weeks. However, OpenAI explicitly noted that this level of government vetting should not become the long-term norm for AI deployment.
How do these tools change the cybersecurity landscape?
Industry experts describe the current environment as an AI-enabled “cat-and-mouse game” between attackers and defenders. Some believe these tools will eventually favor the defenders by allowing them to fix bugs before software is even released.
Lee Klarich of the cybersecurity firm Palo Alto Networks told NPR that the industry is seeing a wave of companies finding and fixing vulnerabilities more efficiently. Klarich suggested this could lead to a future where software is more secure than ever before. Despite this optimism, experts warn that the arms race between hackers and security teams is unlikely to end soon.
Comparison of Government Actions on AI Models
| Company | Government Action | Primary Justification | Current Status |
|---|---|---|---|
| Anthropic | Export Ban (June) | Circumvention of safety guardrails | Partially lifted for select U.S. firms |
| OpenAI | User Screening | National security vetting | Active for newest model release |
The move toward government-mandated screening and export controls marks a departure from the previously “hands-off” approach to AI development. As models like Mythos demonstrate the ability to compress years of manual security work into weeks, federal oversight is likely to focus on the intersection of AI capability and critical infrastructure protection.