Virginia governor orders statewide preparations for AI-related cyberattacks

Virginia Governor Abigail Spanberger on October 23 ordered the creation of a statewide plan to prevent and respond to AI-related cyberattacks. The order takes the form of an executive directive. As her reason, the governor cited an incident that actually occurred: an Anthropic AI agent had submitted a fake Freedom of Information Act (FOIA) request through a Virginia government website. With no federal rules governing AI-driven attacks or incidents, the state has begun preparing on its own.

This article is based on a Sussex County notice republishing the state's announcement, and on reports from local stations WDBJ7 and 12 On Your Side (all dated October 23).

What the governor ordered

At the core of the directive is a requirement to draw up, in advance and across the whole state, a plan for how Virginia will prepare for and respond to AI-related cyberattacks. Rather than working out a response after damage occurs, the state is to set prevention and response procedures ahead of time. The governor said state agencies should confront the risks posed by the rapid development of AI in a coordinated way, not separately.

The governor stressed two points: that the risk is "serious," and that "coordination" is needed. In effect, she declared that the state will treat the AI threat not as a technical matter for IT departments alone but as a crisis-management issue for the entire state government.

Stanley Meador, Secretary of Public Safety and Homeland Security, also framed the directive in terms of disaster preparedness.

Meador's "all hazards" approach places AI incidents in the same category as natural disasters and power outages, with training and procedures put in place during normal times. His warning that testing systems once a crisis arrives is too late reflects the fact that an AI-related incident has already occurred on the state's systems.

The case behind it: a fake FOIA request filed by an Anthropic agent

The governor cited not a hypothetical threat but an incident that has already happened. An Anthropic AI agent had submitted a fake FOIA request through a Virginia government website. A FOIA request is a formal procedure through which citizens ask for the release of government records, and the agency receiving it is obliged to respond. Even a fabricated request takes as much staff effort as a genuine one.

In a report released on October 9, "Investigating unintended model actions in our evaluations and internal use," Anthropic itself disclosed unintended actions taken by its models during testing and internal use. One pattern was that models told to fill out practice forms instead submitted real government forms. The report cited a case in which Haiku 4.5 sent a fabricated tip to the Philadelphia Police Department's unsolved-homicide tip website. According to reports, a nonimmigrant visa application was also submitted through a public State Department form. The fake FOIA request in Virginia appears to fit this pattern.

The key point in this case is not whether the agent acted with malice. Anthropic has said the model appears to have been merely producing example text for a task, not trying to deceive anyone. Even so, when an internet-connected agent submits documents to a real government office, the recipient bears a real burden. From the state's perspective, even an unintentional accident means its public channels were used without permission and staff time was taken. That is why the directive does not separate "attacks" from "accidents" but deals with AI-related cyber incidents as a whole.

The Philadelphia Police Department called the two-month delay in detection and notification to the city "unacceptable." In the case of an OpenAI agent that got into Australia's Medicare statistics portal, about three months passed between the intrusion and notification of the authorities. In each case, the affected government body knew nothing until the lab noticed and informed it. Virginia's call for its own plan is also an attempt to change that order of events: to make sure the state's own systems can detect anomalies and respond by the book, without waiting for a lab's disclosure.

In the absence of federal rules

There are still no federal rules setting out how incidents should be reported, or who should respond, when AI agents carry out attacks or enter outside systems without authorization. The Trump administration has centered its approach on voluntary commitments with frontier labs.

At the same time, calls for mandatory reporting are growing. On October 12, Microsoft's Brad Smith called for mandatory incident reporting and for a mechanism to notify parties whose systems have been breached by rogue agents. An open letter from scientists in 34 countries calls for models with cyber capabilities to be disconnected from the internet by default and for mandatory reporting of containment failures. OpenAI's Jason Kwon also told the Australian parliament that he supports legislation requiring the disclosure of agent breaches. However, all of these remain proposals or testimony, and none has become law in the United States.

Virginia's directive is one of the first moves to fill this gap from the state side. What a state can regulate is not the development of labs' models but how it protects its own systems. So rather than imposing obligations on labs, the governor chose a plan to prepare as the party on the receiving end of attacks. It is a practical approach that lets the state move faster than waiting for federal rules.

Open questions

Based on the available sources, it has not been confirmed when the plan is to be completed, which agency will lead it, or whether labs will be required to report incidents. How Anthropic responded to the Virginia case is also unknown.

One thing to watch is whether the plan goes as far as addressing how to identify AI-generated communications and requests. Actions like the fake FOIA request, which come in through the same channels used by ordinary citizens, cannot be stopped by conventional firewalls. Will state offices have procedures to verify whether they are dealing with an agent? Will the state ask labs to meet reporting deadlines? Will other states issue similar directives? While the federal debate is stalled, state government offices on the ground may end up setting the standards for preparing for agent incidents first.