Congress is no longer treating frontier AI cyber risk as a lab-only problem. Axios reported on April 28 that OpenAI and Anthropic separately briefed House Homeland Security Committee staff on newly cyber-capable models and what those systems could mean for critical infrastructure defense.

The specific model details matter. Axios says Anthropic has held back a public release of Mythos Preview because it can quickly find and exploit critical security flaws. OpenAI, meanwhile, chose a tiered release path for GPT-5.4-Cyber. That is a useful signal on its own: both labs appear to believe the biggest near-term policy risk is not abstract superintelligence rhetoric but operational abuse of models that can accelerate reconnaissance, vulnerability discovery, and exploit development.

Axios also reports that the briefings were classified and took place on Thursday, according to a committee aide. OpenAI said it held several briefings with House and Senate committees and with the White House last week. Anthropic said congressional staff are regularly briefed on model capabilities and their national-security implications. The discussions also touched on a recent White House memo accusing China of industrial-scale efforts to distill and copy American AI models, underscoring how cyber capability and model security are now converging into the same policy lane.

The political backdrop is catching up fast. House Homeland Security Chair Andrew Garbarino framed industry-government coordination as essential for both defensive use and protecting U.S. AI development. At the same time, lawmakers who saw separate demos of jailbroken systems came away describing the capabilities as frightening and Congress as far behind the technology. That reaction matters because future AI legislation may be driven less by consumer chatbot complaints and more by national-security committees looking at concrete misuse paths.

The bigger shift is psychological. Once lawmakers see frontier models as tools that can pressure hospitals, utilities, pipelines, and other thinly staffed infrastructure operators, AI governance stops looking like a distant ethics seminar. It becomes a homeland security file. That frame could shape how future access tiers, red-teaming requirements, and disclosure rules are written. The source article is here.

#congress

OpenAI and Anthropic take cyber-capable models to Capitol Hill