Mary Jane Wilson-Bilik
Partner

Mary Jane (MJ) Wilson-Bilik is a technology, privacy and insurance regulatory law partner at Eversheds Sutherland (US) in Washington, DC. She serves as head of the firm’s US artificial intelligence (AI) practice in financial services, and co-chair of its Global AI in financial services practice. For more than 25 years, MJ has advised her financial services and technology clients on privacy and innovation issues.

Her recent focus has been to assist clients in performing AI impact assessments and develop robust risk management, governance and vendor management policies for their AI efforts. She counsels clients on assessing AI risks on a cross-sector basis; monitoring for bias, privacy, cybersecurity, intellectual property and specialized regulatory risks; and implementing effective governance measures and guardrails. MJ is a frequent speaker at conferences and client meetings on privacy, risk management and artificial intelligence. MJ received her PhD in quantitative political science from Columbia University and her J.D., magna cum laude, from the Georgetown University Law Center.

KEY AI DEVELOPMENTS AND TRENDS IN THE US FROM A LEGAL PERSPECTIVE

It has been a year since the Trump Administration in the United States published “Winning the Race: America’s AI Action Plan,” (“AI Action Plan” or “Plan”), which called for promoting US dominance in AI on the world’s stage, advancing US AI innovation and infrastructure with few, if any guardrails, addressing cybersecurity threats involving AI, and promoting exports of US-developed AI technology¹“The Trump Administration’s plan to win the AI race – a legal perspective,” Eversheds Sutherland Legal Alert (Aug. 6, 2025).. Recently, advanced frontier AI models have displayed extraordinary capabilities to identify new cybersecurity vulnerabilities in systems once thought safe, and AI Agents have gone rogue in varied ways that experts call concerning. This article provides an update on US AI policy and legal developments since the Plan was announced, with a focus on government and industry responses to the wake-up call posed by the enhanced abilities of frontier AI models, the proliferation of rogue autonomous AI Agents, the efforts by the Trump administration to preempt state AI laws and the call for investigations at all levels of government into the AI ecosystem.

The AI Boom in the US

AI has become a huge part of the US economy since the AI Action Plan was issued, with the private sector spending hundreds of billions of dollars to meet AI computing needs. In addition, the Commerce Department recently reported that construction outlays on data centers reached $68.3 billion in the past year.²“The AI Boom is Transforming the American Economy Beyond Recognition,” The Wall Street Journal, by Justin Lahart (Aug. 3, 2026). Both metrics are just a few of the measures that demonstrate how AI is increasingly an integral part of the US economy and lifestyle, as had been anticipated in the AI Action Plan. But along with the economic boom comes legal risks, policy announcements and political debates on many fronts fueled in part by the fast pace of change and uncertainty about where the AI boom is heading.

Frontier AI Models: New Cybersecurity Capabilities

The past year saw the continued rapid development of more and more powerful and sophisticated AI models – called frontier models – by Big Tech that performed feats that both amazed and disturbed experts and policymakers alike. A frontier model is a highly capable AI model that represents the state of the art at a given time and whose capabilities are significant enough to attract heightened attention from regulators, policymakers and AI safety researchers. Such capabilities include automated hacking and sophisticated tactics.

In April and May 2026, several frontier models were premiered that demonstrated unique abilities to find and exploit previously undetected cybersecurity weaknesses in widely used computer systems and software in the financial services sector and throughout the economy. This convergence of AI and cybersecurity has sparked calls to action across sectors, especially financial services. Large cloud computing companies and financial institutions began working with the new frontier models to find vulnerabilities in their own systems, uncovering flaws at a scale and depth in legacy systems that was previously unheard of. For example, a 27-year-old vulnerability was discovered in an operating system widely used in firewalls and critical infrastructure and a 17-year-old vulnerability was discovered that allowed an unauthenticated attacker anywhere on the internet to gain root access to a server. The discovery of these weaknesses in systems previously thought to be secure has led to recommendations that frontier AI models be used in cybersecurity vulnerability management.³“AI-based vulnerability testing to compete in the digital arms race,” Eversheds Sutherland Legal Alert (Aug. 5, 2026). See this alert for a full discussion of using AI-based vulnerability testing.

International and US Governmental Reactions

In response to revelations of the extraordinary capabilities of new frontier AI models to detect new system vulnerabilities, the US governmental, state and international organizations have issued directives and calls to action.

In May 2026, the International Monetary Fund (IMF) warned that AI is “transforming how the financial system copes with vulnerabilities and reacts to incidents.” ⁴“Financial Stability Risks Mount as Artificial Intelligence Fuels Cyberattacks,” IMF Blog (May 7, 2026). The IMF report characterized new frontier AI systems as amplifying cyber threats that can undermine financial stability because the offensive capabilities of intruders could outpace defenses. This is of particular concern in modern international financial systems, which are built on common software and shared service providers that could create simultaneous vulnerabilities across many institutions. The IMF raised the specter of a potential macro-financial shock where payment disruptions, liquidity strains and fire-sale dynamics could occur if multiple institutions are affected simultaneously. The IMF has called for immediate action.

On June 2, 2026, President Trump signed new Executive Order (EO) 14409 that recalibrates aspects of the laissez-faire approach to AI oversight articulated in the AI Action Plan by setting out steps for the federal government to collaborate with industry to strengthen cybersecurity protections in light of the new risks presented by advanced frontier AI models.⁵Promoting Advanced Artificial Intelligence Innovation and Security,” Executive Order 14409, (June 2, 2026). The EO establishes a framework that allows developers of frontier AI models to submit their models voluntarily to the US federal government within 30 days prior to public release to enable testing for cybersecurity risks by the US government. The EO directs the federal government to establish a classified benchmarking review process to determine which models are “covered frontier models” and to assess their cyber capabilities. The benchmarks and review process are highly secretive, so little is known about the details of the process, which is controlled by the Director of the National Security Agency (NSA). Voluntary participants will enter into protected agreements with the government that provide certain confidentiality, cybersecurity, insider risk and intellectual property protections.

Second, the June 2nd EO creates a voluntary “clearinghouse” formed by the Treasury Department and other governmental agencies to identify and address security vulnerabilities in critical infrastructure. On July 14, 2026, the White House announced the launch of “GOLD EAGLE,” the new public-private clearinghouse that operationalizes the EO by using frontier AI models to ingest and prioritize vulnerabilities in company systems at scale and provide a mechanism for remediation and disclosure of prioritized threats and remedial measures to both government and the private sector.⁶White House Launches Gold Eagle Initiative for Unprecedented Cybersecurity Vulnerability Coordination,” The White House” (July 14, 2026). Companies voluntarily participating in the clearinghouse should carefully assess the legal protections being provided for the information they submit to the clearinghouse. Given the Treasury Secretary’s role under the EO in taking the lead on the clearinghouse, financial institutions should expect early outreach from federal authorities to participate in GOLD EAGLE and its vulnerability scanning and remediation potential for systems supporting the financial sector.

On June 22, 2026, the international consortium of cybersecurity agencies known as Five Eyes – representing the United States, United Kingdom, Canada, Australia and New Zealand – issued a statement calling for action now to improve cyber defenses as frontier AI models are exceeding anticipated offensive and defensive cyber capabilities and lowering the barriers for malicious actors to act with speed and complexity.⁷Five Eyes Cyber Security Agencies Statement,” CISA (June 22, 2026).

On the state level, New York’s Department of Financial Services (NY DFS) issued an Advisory on May 21, 2026, about the heightened cybersecurity risks that frontier AI models pose for financial institutions doing business in New York.⁸Heightened Cybersecurity Risks Associated with Frontier AI Models,” Department of Financial Services (May 21, 2026). The Advisory urged all individuals and entities regulated by the NY DFS to expedite their vulnerability management capabilities and coordinate with third-party service providers regarding specific responsibilities for assessing and remediating vulnerabilities. It also recommended that the entities using AI to identify and remediate cyber weaknesses employ secure programming practices to prevent unknown changes in code or configurations, or inadvertent destruction or degrading of necessary code.

AI Agents

Along with rapid change to frontier AI models’ ability to detect unknown cybersecurity vulnerabilities, this past year has seen the proliferation of sophisticated AI Agents that can autonomously pursue goals, make decisions and take actions, often across multiple business units and organizations, without human input. AI Agents operate in self-directed loops. They plan, act, observe results, adjust and repeat their actions with persistence until the task is done. AI Agents vary in the extent to which humans are in control, whether the Agents’ actions align with human authority and expectations, and whether they are transparent and protect privacy and security. AI Agents do not just assist human decision making; they make and execute decisions independently, performing tasks in seconds that would take humans days or weeks to accomplish.

In July and August 2026, the focus on AI Agents intensified as it was revealed that several AI Agents in high-stakes simulations in both the US and the UK went “rogue,” taking unauthorized actions in the real world and behaving in ways that were deceptive and/or potentially harmful. Several major developers of AI Agents reported that their AI Agents had escaped their testing environments, accessed the internet and in some cases hacked into one or more unsuspecting third-party systems during a cybersecurity challenge. Some experts expect more such incidents to occur in the future that may bring additional attention to AI Agent safety and the potential national security issues these incidents may represent.

Research has shown that autonomous AI Agents are most likely to misbehave when a model anticipates that it is being replaced, shut down or restricted. In such cases, it may take preemptive actions to avoid these outcomes that exceed its authorized boundaries. Also, when the model’s assigned objectives conflict with changes in organizational strategy or external constraints, it may act in ways that preserve its original goals, even if its behavior is harmful. Research emphasizes the importance of careful deployment, continuous monitoring and further study of AI alignment to prevent potential harm in increasingly autonomous systems.

Financial regulators are taking note of the risks posed by AI Agents. For instance, the Financial Industry Regulatory Authority (FINRA) that regulates broker-dealers in the US has expressed concern that AI Agents may act beyond a user’s intended authority, may inadvertently expose or misuse sensitive data and may have difficulty tracing the multiple steps it took when making decisions.⁹“FINRA flags rise of agentic AI, seeks member firms’ feedback,” InvestmentNews, by L. Almazora, (Jan. 27, 2026). FINRA has called for robust supervision, clear limits on the scope and authority of AI Agents, and strong logging and audit capabilities. For now, as with most US financial regulators, FINRA is relying on its existing regulatory framework of disclosure, fiduciary duty, supervision, vendor management and cybersecurity to regulate any AI-driven incidents that occur.

Preemption of State AI Laws by Federal Authorities

In response to the AI boom, many states have enacted new laws and regulations that require developers and/or deployers of AI to enact guardrails around the operation of AI systems, some of which have been viewed by the current Administration as excessive. In response, President Trump signed an Executive Order, “Ensuring a National Policy Framework for Artificial Intelligence”¹⁰Ensuring a National Policy Framework for Artificial Intelligence,” Executive Order 14365 (Dec. 11, 2025). on December 11, 2025 (2025 EO), that orders federal agencies to take steps to evaluate and challenge on constitutional grounds state AI laws that the Administration deems excessive, while calling for federal legislation and federal agency actions that could preempt such state AI laws.

The 2025 EO asserts that certain state laws require companies to “embed ideological bias within models,” and cites to Colorado’s former AI law, SB24-205 that bans algorithmic discrimination, as an example of an excessive state law that may require AI models to produce false results in order to avoid a “differential treatment or impact” on protected groups. This provision of the 2025 EO echoes Executive Order 14319, “Preventing Woke AI in the Federal Government” (July 23, 2025)¹¹“Preventing Woke AI in the Federal Government,” Executive Order 14319 (July 23, 2025). For further information, see “The Trump Administration’s plan to win the AI race – a legal perspective,” Eversheds Sutherland Legal Alert (Aug 6, 2025). that requires all Large Language Models (LLMs) procured by the federal government to produce reliable outputs free from harmful ideological biases and social agendas.

The 2025 EO is not self-executing. Rather, it lays the groundwork for federal challenges to state AI laws, for new regulations and federal laws that would preempt state AI laws, and for restrictions on federal funding to states with “onerous” AI laws, as identified by a joint Department of Commerce/White House Task Force. To date, while no actions have been taken by the Trump Administration to operationalize the 2025 EO, other than the creation of an AI Litigation Task Force at the Department of Justice, state officials and various interest groups remain alert to any developments that may infringe on the states’ perceived right to regulate AI.

No Uniform AI Rulebook but Investigations under Existing Laws

While the contours of a comprehensive national law to regulate AI have eluded lawmakers to date, federal agencies, state attorneys general, state regulators and congressional committees are nonetheless undertaking investigations into the AI ecosystems using existing legal authorities as their guideposts.

The Federal Trade Commission (FTC) on July 7, 2026, proposed a policy statement clarifying that companies that develop or deploy AI systems (AI companies) remain subject to the FTC Act’s prohibition on unfair or deceptive acts or practices. The proposed policy statement states that AI companies may not deceive consumers by manipulating the AI system’s behavior in ways that are contrary to the reasonable expectations of consumers regarding objectivity and accuracy. AI companies should expect FTC regulators to compare what a company says publicly with what its internal communications and testing show privately.¹²“FTC’s policy statement on ‘suppression of accuracy’ in AI systems takes aim at states’ efforts to regulate AI: What it means for businesses,” Eversheds Sutherland Legal Alert (July 9, 2026).

State attorneys general are also playing a major role in AI regulation. In August 2025, a bipartisan coalition of 44 state attorneys general wrote to major AI companies regarding child safety and chatbot interactions.¹³“Bipartisan Coalition of State Attorneys General Issues Letter to AI Industry Leaders on Child Safety,” National Association of Attorneys General, (Aug. 26, 2025). While the August 2025 letter does not propose specific investigatory actions, it serves as a transparent and proactive notification to the AI industry that the coalition is vigilant to ensuring children remain safe from harm and exploitation.

State insurance regulators are undertaking examinations into AI systems used by insurance companies in their operations to determine whether insurers have sufficient governance, risk management and vendor management controls in place to ensure compliance with state insurance laws. In July 2025, the National Association of Insurance Commissioners’ Big Data and Artificial Intelligence (H) Working Group released for comment a draft artificial intelligence systems evaluation tool. That tool is undergoing field testing in targeted examinations and is expected to be finalized by the end of 2026.¹⁴“A Look At NAIC’s Proposed Tool For Evaluation Of Insurer AI,” Law360 (Aug. 8, 2025).

At least seven Congressional Committees are investigating various aspects of the AI landscape, including investigations into recent autonomous AI Agents hacking incidents. In April 2026, the House Select Committee on the Chinese Communist Party (CCP), together with the Homeland Security Committee, announced a formal investigation into national security risks posed by CCP-developed AI models. The committees sent inquiry letters to US companies, including companies deploying Moonshot AI’s Kimi model. Also in April 2026, several developers of frontier AI models briefed staff of the House Committee on Homeland Security regarding increasingly capable cyber-focused AI models and the risks they may pose. The committee has been examining how advanced models could be used offensively and how critical infrastructure operators should prepare.

Investigations by federal, congressional, and state officials into the practices of AI developers and deployers are expected to continue as powerful frontier AI models and autonomous AI agents become more capable and increasingly outpace human workflows. The regulatory trajectory suggests that regulators and investigators alike will focus closely on who will be liable for any harm caused by autonomous AI.

Conclusion

As AI systems and AI Agents are increasingly integrated into commercial and government applications, there is a growing demand to govern, manage and monitor these systems and their outputs in real time. While the concept of monitoring AI systems for legal compliance is not new, best practices for governing and managing AI risks are still evolving and require constant vigilance. Exactly because AI systems are rapidly evolving in ways that introduce new risks and unpredictability into their ecosystems, strong governance practices, AI-based vulnerability testing, risk assessments and post-deployment monitoring at machine speed are crucial practices for confident, wide-spread AI adoption.