Anthropic Mythos fears push Google, Microsoft and Elon Musk’s xAI into new US government AI safety pact

According to details shared by the Department of Commerce, CAISI will conduct pre deployment evaluations and specialized research on frontier AI models developed by participating companies. These assessments are designed to better understand the capabilities, limitations and security risks associated with next generation AI systems.

The initiative expands earlier partnerships established in 2024 with companies such as OpenAI and Anthropic. Officials say the latest agreements strengthen the government’s ability to independently evaluate powerful AI technologies before they become widely accessible.

In its official statement, the Department of Commerce said the expanded collaborations are intended to improve AI security research while aligning with the broader goals of America’s AI Action Plan.

The statement added that the agreements were renegotiated to reflect updated directives from the secretary of commerce and evolving national priorities surrounding advanced AI systems.

At the center of the initiative is the growing belief within Washington that highly capable AI models may eventually create risks that extend far beyond ordinary consumer technology. Government officials increasingly view frontier AI as a strategic technology with implications for national defense, economic competition and cyber resilience.

The immediate backdrop to the new agreement is the emergence of Anthropic’s Mythos system, which reportedly heightened concerns about how advanced AI tools could be misused.

While officials have not publicly disclosed technical details about the model, reports suggest that policymakers became concerned over AI systems that may assist hackers in identifying vulnerabilities, generating malicious code or accelerating cyberattack planning.

The fears are not limited to criminal hacking alone. Security experts have warned that powerful AI systems could eventually help hostile actors automate cyber intrusions, spread disinformation campaigns or support military intelligence operations.

These concerns have pushed governments worldwide to reconsider how AI systems should be evaluated before deployment. The United States, which remains home to many of the world’s leading AI companies, is now attempting to balance innovation leadership with stronger oversight mechanisms.

The inclusion of companies such as Google DeepMind, Microsoft and xAI demonstrates that even the industry’s biggest players recognize the growing pressure for accountability and independent testing.

Under the arrangement reported by Reuters, the US government will receive a “first look” at some of the participating companies’ most advanced AI models before they reach the public.

This process is expected to include several layers of testing and evaluation.

Government researchers will assess whether frontier AI systems could enable cyber warfare related activities or create risks tied to military misuse. Officials will also study the overall capabilities of these systems to understand how powerful they have become and where safeguards may still be weak.

One of the most important aspects of the evaluation process involves testing versions of the AI software where certain safety protections or operational restrictions have been removed. By analyzing how the underlying models behave without their standard safety layers, researchers hope to better understand the technology’s raw capabilities and potential vulnerabilities.

The goal is not only to identify immediate dangers but also to establish scientific standards for evaluating future AI systems as the technology rapidly evolves.

Chris Fall, Director of CAISI, stressed the importance of independent scientific evaluation in understanding the broader national security implications of frontier AI.

He said rigorous measurement science is essential for determining the real world risks associated with advanced AI models and for ensuring that evaluations are conducted in the public interest.

The agreement also reflects a broader shift in how governments now view artificial intelligence. Just a few years ago, AI discussions were largely focused on productivity, automation and consumer applications. Today, advanced AI systems are increasingly being treated as strategic assets with direct implications for national power.

Countries around the world are investing heavily in AI infrastructure, semiconductor supply chains and advanced research capabilities. At the same time, policymakers are becoming more aware that frontier AI models could potentially be weaponized if left unchecked.

The United States has been especially active in trying to shape AI governance while maintaining its competitive edge against global rivals. Washington has introduced export controls on advanced chips, increased scrutiny of foreign AI investments and encouraged closer cooperation between government agencies and private AI companies.

The new CAISI agreements fit into that larger strategy by creating a structured channel through which government experts can study frontier AI systems before public deployment.

Industry leaders appear increasingly willing to participate in such initiatives, partly because public trust in AI safety is becoming essential for long term adoption.

The participation of Google DeepMind, Microsoft and Elon Musk’s xAI carries significant weight because all three companies are heavily involved in the development of large scale AI systems.

Google DeepMind remains one of the world’s most influential AI research organizations and continues to compete aggressively in the generative AI race. Microsoft, through its deep partnership with OpenAI and its broader AI investments, has become a dominant force in enterprise AI infrastructure. Elon Musk’s xAI, meanwhile, has rapidly expanded its ambitions as Musk positions the company as a major challenger in the global AI market.

By agreeing to government evaluations before public releases, these companies are signaling that AI safety concerns can no longer be treated as secondary issues.

The move may also help the companies demonstrate responsible development practices at a time when regulators worldwide are demanding greater transparency from AI developers.

Public scrutiny of artificial intelligence has intensified over concerns involving misinformation, deepfakes, copyright disputes, bias, surveillance and cybersecurity threats. Governments are now under pressure to ensure that advanced AI systems do not outpace existing safeguards.

The agreements with CAISI could become an early model for how future AI oversight frameworks operate in the United States.

The Center for AI Standards and Innovation has rapidly become one of the most important government institutions involved in AI evaluation and safety testing.

According to the Department of Commerce, CAISI has already completed more than 40 evaluations of AI models, including assessments involving systems that have not yet been publicly released.

That growing body of work is helping position the agency as a central hub for independent AI testing in the United States.

Officials believe the organization can provide scientific credibility and technical expertise at a time when governments need more reliable ways to evaluate rapidly evolving AI systems.

The agency’s work is also expected to influence future policy decisions, especially as lawmakers debate how to regulate frontier AI technologies without slowing innovation.

Industry experts say independent evaluations may eventually become a standard part of AI development, much like safety testing in industries such as aviation, pharmaceuticals and automotive manufacturing.

The latest agreement between major AI companies and the US government underscores a growing reality within the technology sector. Artificial intelligence is no longer viewed only as a commercial breakthrough. It is increasingly seen as a technology capable of reshaping cybersecurity, military strategy, economic competition and global influence.

As AI models become more capable, governments are likely to demand stronger oversight, deeper transparency and more rigorous safety evaluations.

At the same time, technology companies remain under pressure to innovate quickly in one of the most competitive industries in the world.

The challenge for policymakers and AI developers alike will be finding a balance between encouraging technological progress and preventing dangerous misuse.

The new CAISI agreements represent an important step in that direction. Whether they become a long term model for AI governance may depend on how effectively these evaluations identify risks while still allowing innovation to move forward.

For now, the agreement signals that the era of unchecked AI deployment is rapidly changing, and that national security considerations are becoming central to the future of artificial intelligence.