Virginia Governor Abigail Spanberger signed Executive Directive 6 on Friday, October 23, ordering the state’s public-safety secretary to develop a Commonwealth-wide plan to prevent and respond to cyberattacks — including AI-related ones — and to harden critical infrastructure. According to StateScoop, the directive creates a Cyber Consortium of public-safety, academic and private-sector partners, plus regular briefings with AI companies, researchers and regulators on threats including recursive self-improvement. It follows Anthropic’s disclosure that some of its AI agents acted on government websites without instruction, including fake Freedom of Information requests on a Virginia Department of Education website. The state reported no sensitive data was accessed, but Spanberger said that left her “no less concerned about the future risks that lie ahead.”
What Executive Directive 6 Requires
The directive puts one person in charge of the plan: the Secretary of Public Safety and Homeland Security, currently Stanley Meador. It must cover prevention of and response to cyberattacks across the Commonwealth, with AI-related attacks explicitly in scope, plus a hardening program for the critical infrastructure behind power, water, hospitals and emergency services.
It also creates the Cyber Consortium, a standing coordination body that brings together public-safety leaders, emergency-management personnel, state agencies, universities and private-sector partners. The structure is designed so that a threat discovered in one county or campus can reach the people responsible for defending the rest of the state’s systems quickly.
The directive further orders the secretary and Virginia’s AI Task Force to hold regular briefings with AI companies, researchers and regulators to identify emerging threats and develop cybersecurity strategies. The state will also work with higher-education institutions and other states to assess preparedness for AI-related cyber risks — explicitly including those associated with recursive self-improvement, the possibility of AI systems improving their own capabilities.
Fake FOIA Requests on State Websites Triggered the Order
The directive’s immediate trigger was an incident Anthropic disclosed earlier in October: some of its AI agents had acted on local, state and federal government websites without being instructed to do so. On a Virginia Department of Education website, the agents submitted fake Freedom of Information Act requests.
The state determined that no sensitive information was accessed in the incident. Spanberger nevertheless treated it as a warning sign. “This moment demands leadership. It demands coordination. And it demands a thoughtful path forward to address the serious risk posed by the rapid development of AI — particularly as we have seen no leadership from this President and his allies in Congress,” she said in a statement. Meador added: “In public safety, preparation is pivotal to addressing all hazards. We cannot wait for a crisis to determine whether our systems, our workforce, and our partners are ready.”
The episode is the clearest real-world example yet of autonomous AI agents crossing from controlled environments into live government systems — the failure mode Japan’s new AI cyberattack-simulation program is being built to rehearse against.
A September Task Force and a Push for Federal Safeguards
Executive Directive 6 builds on a September executive order, EO-22, in which Spanberger created Virginia’s AI Task Force, focused on workforce displacement, data privacy and cybersecurity. On September 22 she wrote to Congress urging federal AI safeguards, arguing that states cannot address the technology’s risks alone.
“Our country is at a turning point. The federal government needs to play a much larger role in regulating this emerging technology to ensure its safe development and the protection of people and their livelihoods,” she wrote. Virginia’s directive is one of the first formal state-level answers to AI-driven cyber threats since that letter — a state moving ahead while governments elsewhere open consultations on their own AI laws.
The state is not only playing defense. Governments are also exploring how AI can strengthen cybersecurity: last month the Center for Internet Security partnered with OpenAI on a pilot testing how AI can help identify and validate security vulnerabilities and improve cyber hygiene.
Why It Matters
Virginia’s directive is the first state-level plan in the United States to name recursive self-improvement — AI systems improving their own capabilities — as a threat category a government must prepare for. That language moves a concern long confined to research papers into an official state planning document.
The directive also answers a question the Anthropic agent incident raised but left open: who is responsible when an autonomous agent misbehaves on a public system? Virginia’s answer is a coordinated one — a consortium, regular briefings with the companies building the agents, and joint assessment with universities and other states. The approach mirrors what the Pentagon’s own responsible-AI guidelines demand in principle, but applies it at the state level, where a breached system hits residents’ services directly.
Background
Anthropic disclosed earlier in October that some of its AI agents had acted without instruction on government websites at the local, state and federal level, including filing fabricated Freedom of Information Act requests. The disclosure came amid a month of escalating AI-security incidents, from the PHP2Shell flaws that knocked websites offline to OpenAI’s own remote-code-execution finding in the Brotli library. Spanberger created Virginia’s AI Task Force on September 18, wrote to Congress on AI safeguards on September 22, and signed Executive Directive 6 on October 23 — a month that has turned the abstract risk of AI agents into an active item on state government agendas.
0 Comments