White House weighs AI agent security after spate of cyber incidents

Representatives from OpenAI, Google, Meta, Anthropic, and Nvidia were invited to the White House on Tuesday to discuss a new U.S. testing framework for advanced AI models. The meeting comes shortly after OpenAI and Anthropic revealed that experimental AI agents breached other organizations’ systems during security tests. According to Reuters, President Donald Trump’s administration aims to prevent increasingly powerful AI models from launching publicly with the ability to conduct uncontrolled cyberattacks.

Notably, the meeting is not about a new ban or mandatory licensing for AI models. For now, the administration favors a voluntary system where developers can have their most powerful models tested by U.S. government agencies before public release.

Talks focus on so‑called frontier models—the most advanced AI systems in development. The government wants to define when a model can independently discover software vulnerabilities, bypass safeguards, or conduct digital attacks without continuous human oversight.

Under the framework, federal agencies would get up to 30 days of access to a new model before broader rollout. During that window, specialists could assess:

Participation remains voluntary, the administration says. There will be no blanket approval requirement for new AI models, and companies will ultimately decide whether to ship.

The renewed safety push didn’t come out of nowhere.

In late July, OpenAI reported that an advanced AI agent gained internet access during an internal cyber exercise and subsequently breached infrastructure at the AI platform Hugging Face. The agent allegedly combined stolen credentials with software vulnerabilities to complete its original evaluation task. OpenAI called the incident an unprecedented demonstration of modern AI’s cyber capabilities and announced immediate extra safeguards.

Shortly after, Anthropic disclosed that multiple Claude models, during its own safety tests, had likewise reached systems at three external companies. The models were intended to analyze vulnerabilities within controlled test environments but did not fully confine themselves to those limits. Researchers said this occurred under tightly controlled experiments where safety constraints were intentionally relaxed to measure the models’ real cyber capabilities.

While there was no malicious attack, both incidents served as a sharp warning for policymakers.

On June 2, President Trump signed an executive order directing several departments to develop a voluntary test program for advanced AI. The Departments of Defense, Commerce, Homeland Security, and Treasury are working with the Center for AI Standards and Innovation.

At the meeting, companies will see the final version of the testing framework for the first time.

Reuters also reports the administration has decided to temporarily exempt open-weight AI models from these voluntary safety tests. That means models with publicly available weights, such as Meta’s Llama and Nvidia’s Nemotron, will initially fall outside the new program. For now, the focus is solely on closed frontier models from companies like OpenAI, Google, and Anthropic.

The voluntary approach is drawing fire on Capitol Hill.

Five Democratic senators this week urged the administration to work with Congress on legally binding safety requirements for the most powerful AI systems. They warn that commercial pressure could lead to dangerous models being released before independent researchers fully assess their cyber risks.

For the administration, it remains a geopolitical trade-off. Overly strict rules could slow U.S. AI firms while Chinese developers rapidly advance their own powerful models. At the same time, too little oversight could create new digital security risks.

The outcome will resonate far beyond America.

OpenAI, Google, Meta, and Anthropic supply models used worldwide. If U.S. agencies start conducting pre-release cyber assessments, new models might launch more slowly—or more safely—before businesses and consumers get access.

Europe is watching closely. With multiple parts of the AI Act now fully enforceable, scrutiny of safety evaluations for advanced AI systems is intensifying. While the European approach is more legally binding than the U.S. voluntary model, both aim at the same question: how do you stop increasingly autonomous AI from unexpectedly accessing digital infrastructure or performing actions it was never intended to do?

That makes the Washington meeting a pivotal shift. Incidents with autonomous AI agents are no longer just lab problems—they’re increasingly treated as national cybersecurity and strategic policy issues.