Congress Targets AI Model Theft With the Bipartisan AI Cyber Defense Act

Two members of the US House of Representatives from opposite parties have introduced a bill aimed at one of the quieter threats in artificial intelligence: the copying of American AI models by foreign actors. The proposal, called the AI Cyber Defense Act, would put the National Institute of Standards and Technology, known as NIST, at the center of the response. It does not ban any technology outright. Instead, it tries to build a shared defense, with voluntary standards and a standing table where government and industry can compare notes.

What the bill would do

According to reporting by Nextgov and CyberScoop, the bill was introduced on October 22, 2026 by Representative Pat Harrigan, a Republican from North Carolina, and Representative George Whitesides, a Democrat from California. Its core task for NIST is to develop voluntary industry standards and best practices for preventing distillation attacks against American made AI models, working in partnership with private sector model developers.

The bill would also create an AI Cybersecurity Roundtable, bringing critical infrastructure operators together with tech firms to develop stronger cyber defenses for AI capabilities. Power grids, hospitals, and financial systems are the examples Harrigan named in a statement. A summary of the sponsor’s press release adds that NIST would have 180 days after enactment to begin engaging stakeholders, and that the resulting standards would be made public.

What a distillation attack actually looks like

Distillation sounds technical, but the idea is simple. A powerful AI model is expensive to build, needing huge computing power, large datasets, and long training runs. A distillation attack tries to skip most of that cost by learning from the finished model instead.

In practice, according to CyberScoop and SC Media, an attacker sends a very large number of prompts to a target model, often through many throwaway or burner accounts, and collects the answers. Those question and answer pairs are then used to train a second model. The copy is rarely identical, but it can get close enough to be useful, and it can be built for a fraction of the original cost. That is why lawmakers describe it as theft, even though it does not involve breaking into a server in the traditional sense.

There is an honest complication. Some forms of distillation are a normal, openly discussed part of AI research, and open weight models are often built, in part, by learning from other models with permission. CyberScoop reports that NIST will first have to define what counts as a distillation attack, and separate hostile copying from the sharing and collaboration that has helped drive innovation. That definition will probably be the hardest part of the whole effort.

Why lawmakers say this moment matters

Harrigan framed the bill as a race against time. In a statement quoted by Nextgov, he said that some of the greatest advances in artificial intelligence are happening in America, and that the country has a responsibility to make sure that technology is secure. He added that adversaries are already looking for ways to exploit these systems, and that as AI becomes more powerful, the risks to critical infrastructure will only grow.

Reporting on the bill names Russia, China, Iran, and North Korea, along with entities tied to them, as actors of particular concern. CyberScoop reports that US officials and companies including Anthropic, OpenAI, and Google have identified systematic campaigns in which actors use burner accounts to extract data from frontier models, the most capable systems available.

The bill also reflects a worry about competitive lead. CyberScoop reports that Harrigan believes the US AI industry holds a lead of roughly six to twelve months over other countries, and that maintaining it is incredibly expensive. In an interview, he argued that distillation gives other parties an incentive to hang back, let American companies pay for the hard research, and then copy the results for a small share of the cost.

The other AI bill from the same week

The distillation bill was not the only AI proposal in Congress that week. Nextgov reports that Representative Shontel Brown, a Democrat from Ohio, introduced a bill on Monday of the same week that seeks to promote transparency around AI generated content in political ads and to curtail uses of AI to affect Americans’ voting rights. One bill protects models from theft, the other protects voters from deception, but they share a premise: AI is now powerful enough that the rules around it can no longer be left entirely to the companies building it. Neither bill is law yet, and both would still need to pass the House, pass the Senate, and be signed by the president.

What the bill does not do, and what happens next

According to the reporting so far, the bill directs NIST to develop voluntary standards, not mandatory rules. Voluntary standards can still shape an industry, because companies often adopt them to show customers they take security seriously, but a company could choose not to follow them without breaking this law. The bill also does not, on the reporting available, spell out how a distillation attack would be detected in practice, or what evidence would prove that one had happened.

The next steps are procedural. The bill will go to committee, where members can amend it, hold hearings, and hear from NIST, AI companies, cybersecurity experts, and infrastructure operators. If it moves forward, the text may change. If it stalls, its ideas may still reappear in other bills.

Frequently Asked Questions

What is the AI Cyber Defense Act?

It is a bipartisan House bill introduced on October 22, 2026 by Representatives Pat Harrigan and George Whitesides. According to reporting by Nextgov and CyberScoop, it would direct NIST to develop voluntary standards against distillation attacks on American AI models and create an AI Cybersecurity Roundtable for government and industry collaboration.

What is a distillation attack?

It is a method in which an attacker sends a very large number of prompts to a powerful AI model, collects the answers, and uses them to train a copycat model. According to CyberScoop, this lets the attacker replicate much of the original model’s capability for a fraction of the original development cost, often using many burner accounts to hide the extraction.

Is the bill already law?

No. It is a proposal that would still need to pass the House and the Senate and be signed by the president. The reporting so far describes voluntary standards, not mandatory rules, and the text could change substantially during committee review.