Law360 (July 6, 2026, 11:29 PM EDT) -- A recent executive order intended to boost the security of advanced artificial intelligence systems hinges on developers voluntarily making their models available to the government for prerelease testing, but lingering questions about the potential trade-offs of this exchange raise doubts about the ultimate effectiveness of this model.

In response to growing national security concerns prompted by the rise of frontier AI models such as Anthropic's Claude Mythos that can rapidly identify and exploit cybersecurity vulnerabilities at scale, President Donald Trump last month issued an executive order that directs several federal agencies to design a framework through which AI developers will be able to voluntarily give the government access to models up to 30 days before their planned release to test for and identify cybersecurity vulnerabilities.

The directive is consistent with the Trump administration's stated desire to promote the deployment of secure models while ensuring that companies still have plenty of room to innovate. But the decision to rely completely on companies to make their models available to the government leaves questions about whether the program can produce sufficient buy-in or beneficial results.

"One thing that jumps out from the executive order is what will the incentives be, and why would companies want to go through with sharing their frontier models with the federal government in the first place?" said ­­­Tyler Thompson, a partner in the emerging technologies practice at Reed Smith LLP.

While the executive order stressed that any sharing of models would be done subject to appropriate confidentiality and intellectual property protections, several potential downsides loom, including uncertainties over what the government will do with and how it will protect from unauthorized disclosure the information it learns from its examination and testing of these models.

"From companies' perspective, it's a little bit of a gamble," said Erin Prest, a partner and co-leader of McCarter & English LLP's cybersecurity and data privacy group. "While voluntarily complying with this executive order and submitting your model to the government ahead of time could garner you some good will or help from the government with finding flaws and vulnerabilities, it may also run the risk of alerting the government to what the model is and how it works."

The latter outcome could be risky in light of the federal government's recent imposition of export controls on a pair of Anthropic's AI models after the company had already worked with the government to test safeguards prior to the technology's release. While Anthropic announced Wednesday that these export controls had been lifted following its further collaboration with the government, AI developers could still encounter delays and other setbacks due to their voluntary disclosures, even though the executive order itself doesn't explicitly bar companies from releasing their products if issues are identified during the testing process, attorneys say.

Additionally, while voluntarily disclosing their models could potentially open doors for developers to participate in other programs and help them avoid surprises that could lead the government to take adverse action against them, there's no guarantee that these companies will get any helpful information back from federal officials in return.

While the executive order directs certain government agencies to develop a "classified benchmarking process" to make assessments about a model's cyber capabilities and its covered "frontier" status that are shared with developers and researchers "as appropriate," companies that share their models may ultimately end up being blocked from receiving any useful feedback due to potential national security risks that could prevent the government from disclosing anything beyond a basic "thanks for sharing" acknowledgment, Prest noted.

"Until there are more operational directives and companies actually start going through the process and seeing what the benefits may be, it's going to be a bit of a mixed bag," noted Prest, who before joining McCarter & English last year served as privacy and civil liberties officer and deputy general counsel at the FBI.

The federal government has turned to such voluntary collaboration in the past to tackle critical cybersecurity threats, with then-President Barack Obama signing an executive order in 2013 to push federal regulators to work with banks, power grid operators and other businesses that operate vital national infrastructure to develop voluntary standards to defend against cyberattacks.

While the initiative produced a voluntary framework to enable critical infrastructure operators to align their cybersecurity protections with common best practices and boost collaboration between the federal government and private sector, Congress moved to give more teeth to this regime through the passage of the Cybersecurity Information Sharing Act in 2015.

That legislation preserved the voluntary nature of the original presidential directive while creating crucial antitrust and legal liability protections to encourage private entities to voluntarily share cyberthreat information among themselves and with the federal government, incentives that are lacking from the latest AI executive order, attorneys noted.

"The information sharing stemming from the [federal legislation] CISA had a very clear carrot for companies in terms of if they shared the threat information, they were protected from lawsuits and other risks on the back end," Prest said. "It's unclear here what the carrot may be [with this executive order]."

While some inherent benefits could emerge for developers that choose to share their models, such as uncovering new capabilities that may allow them to market their models in unexpected ways or getting the opportunity to fix vulnerabilities before they go to market, these benefits are not as explicit as they are in CISA, Prest noted.

Unlike with information-sharing legislation, the new executive order fails to make clear what intellectual property rights developers will retain over the technology they share with the federal government and unspecified "trusted partners" that the directive contemplates will also have access to these models. The presidential directive also falls short in addressing additional issues such as how agencies will view a developer's participation, or lack thereof, when it comes to contracting with the federal government for a wide range of services, noted J.D. Koesters, a shareholder in the government enforcement and investigations group at Baker Donelson Bearman Caldwell & Berkowitz PC.

"There are still unanswered questions about what 'trusted partners' are expected to do and what will be the rights and obligations of developers when they decide to voluntarily submit their models to the program," Koesters said.

Alex Hastings, a partner at Morgan Lewis & Bockius LLP, said it will also be important to see what the government ends up determining the criteria will be for evaluating whether a frontier model qualifies as a national security risk, what happens if that designation is made and what the sharing agreements between AI developers and the government will ultimately look like.

"If there is a template of a model agreement by which companies can share these models with the government that addresses the [outstanding] issues that is fair and balanced, that's going to make this process much easier," Hastings said. "But if there needs to be extensive negotiations between companies and the government every time a model is shared, that's going to make it much more difficult."

Aside from establishing the framework for voluntarily sharing AI models, the executive order also notably directs federal agencies to establish a voluntary clearinghouse to coordinate vulnerability scanning, validation, and patch distribution with AI developers and critical infrastructure operators, another aspect that is poised to have a significant impact on not only participants but also those not directly engaged with the program, noted Stephen Lilley, a partner in the cybersecurity & data privacy and national security practices at Mayer Brown LLP.

"Businesses across the economy depend upon access to patches for vulnerabilities in critical systems," Lilley said. "The clearinghouse appears poised to make very consequential decisions about which systems to test and how to prioritize patch development and distribution to critical infrastructure and other businesses. Monitoring the development and direction of the clearinghouse closely may help companies anticipate implications for their security programs."

Another aspect of the executive order that's drawing less attention, but stands to have a significant impact on both developers and critical infrastructure operators that could benefit from this technology, directs federal agencies to prioritize and expand programs that promote AI-enabled cyber defensive tools and facilitate access to frontier AI models and other cybersecurity services to not only state and local authorities but also critical infrastructure operators such as rural hospitals, community banks and local utilities.

"This part of the executive order could open up a lot of capability and grant funding for AI developers to help critical infrastructure providers that don't have the resources to be integrating these models into the national security space but need the protections," said Hastings, the Morgan Lewis partner, adding that this opportunity presents a chance for developers that have models that could address these cybersecurity concerns in the critical infrastructure space to "be proactive" by offering these capabilities to the government ahead of other developers.

Still, while this collaboration with the government would likely provide a significant boost in the effort to strengthen critical infrastructure providers' cybersecurity posture, the initiative also presents many of the same intellectual property questions and other concerns raised by the sharing of frontier AI models for prerelease testing, making it prudent for developers weighing participation to "consider the tradeoffs and proceed with caution," Hastings said.

Attorneys also pointed out that while the current framework is a voluntary one that reflects the Trump administration's desire to ensure systems are secure without stifling innovation, there's no guarantee that it will stay that way.

"The federal government currently appears to be doing anything it can to enable growth in this space and win the AI race, but the real ask here is that they're hoping that companies will be good corporate citizens about this so they don't have to regulate it more heavily," said Thompson, the Reed Smith partner.

The cybersecurity threat information-sharing regime originally suggested by the 2013 executive order offers a potential parallel. While companies have long been encouraged to disclose details about the vulnerabilities they've encountered and the attacks they've experienced, Congress has been slowly moving to put more pressure on companies to be more forthcoming, first with CISA and then with the 2022 passage of the Cyber Incident Reporting for Critical Infrastructure Act.

The latter statute requires the Cybersecurity and Infrastructure Security Agency to create rules — which the agency said last week that it expects to have finalized by September — to require vital infrastructure operators to report cyber incidents within 72 hours and ransom payments within 24 hours to the agency.

"For a long time, everyone was preaching cybersecurity reporting, but there was really no teeth to that, and it was voluntary until something went wrong, and now we have CIRCIA," said Koesters, of Baker Donelson.

With the potential for AI to be used to orchestrate cyberattacks of unprecedented size and scale, and with states increasingly focusing on this area despite the federal government pushing back against what have so far been primarily mandates for companies to be more transparent about their models' capabilities rather than how they're secured, cybersecurity protocols like those in the latest executive order are likely to remain voluntarily only as long as they're followed, Koesters noted.

"The question is going to be whether the government is going to get ahead of these threats or, as we typically see, if there will be a knee-jerk reaction when something goes wrong," Koesters said. "While we've seen almost nothing but deregulation from this administration on AI, we're seeing some form of ratcheting that regulation back up with this executive order, and that's likely to continue as the government continues to balance their broader strategy with their recognition of the broader implications of these AI models."

--Editing by Jay Jackson Jr. and Emily Kokoll.

For a reprint of this article, please contact reprints@law360.com.