A360
Key Takeaways✦ Atlas AI
01
AI firms are developing cyber-specific models so powerful they are limiting public access and briefing lawmakers on the potential security implications for critical infrastructure.
02
Lawmakers' concerns are escalating, with some describing the technology's potential for misuse as "very scary" and admitting that legislative efforts are "light years behind."
03
The briefings highlight a dual concern: the offensive potential of new AI tools in the hands of adversaries and China's alleged efforts to steal US AI technology.
Atlas AI
Powerful New Models Raise Alarms
Leading artificial intelligence firms have briefed U.S. lawmakers on new models possessing advanced cyber capabilities, highlighting emerging AI cyber threats and accelerating concerns over national security. Staff for the House Homeland Security Committee received separate briefings last week from OpenAI and Anthropic regarding the significant offensive and defensive potential of these new systems.
These discussions, among the first of their kind on such specific tools, underscore a critical juncture where AI development is rapidly outpacing regulatory frameworks. The focus of the meetings was the security implications for the nation, particularly for under-resourced critical infrastructure sectors that could be targeted.
A Cautious Approach to Deployment
In a sign of the technology's perceived power, both companies are taking deliberate steps to manage its release. Anthropic has decided to withhold the public release of its model, dubbed "Mythos Preview," after it demonstrated a high proficiency in identifying and exploiting security vulnerabilities.
Similarly, OpenAI is employing a tiered release strategy for its own specialized model, known as GPT-5.4-Cyber. This measured approach involves collaboration with federal agencies to provide them with controlled access to the tool, ensuring government partners can understand and prepare for its capabilities.
According to a committee aide, the briefings represented "proactive engagement" with the AI labs. A spokesperson for Anthropic confirmed the company regularly informs congressional staff about model capabilities as part of its ongoing national security engagement.
Lawmakers Confront Escalating Risks
The briefings also addressed growing geopolitical tensions, including a recent White House memo detailing accusations of "industrial-scale" campaigns by China to distill and replicate American AI models. This context added another layer of urgency to the conversations on Capitol Hill.
House Homeland Security Chair Andrew Garbarino (R-N.Y.) stressed the need for "productive partnerships between industry and government" to stay ahead of evolving threats. He noted that these briefings help Congress identify risks and develop informed lines of inquiry.
For some lawmakers, the new information compounded existing fears. A prior session demonstrating how "jailbroken" AI models could be manipulated for malicious acts, such as planning attacks, had already created a sense of immediacy. Rep. August Pfluger (R-TX) called that demonstration "very scary," while Rep. Andy Ogles (R-TN) lamented that "Congress is light years behind" the pace of technological change, highlighting the challenge of governing these powerful new tools.
