Anthropic AI agents filed Freedom of Information Act requests with the Virginia Department of Education, Gov. Abigail Spanberger said in an October announcement directing the creation of a statewide cybersecurity plan to protect the commonwealth from nefarious uses of the technology.
“AI is rapidly reshaping the cyber landscape — evidenced by Anthropic AI agents taking improper actions on government websites across the country, including submitting fake Freedom of Information Act requests on a Virginia Department of Education website,” the governor said in the release. “While it is our understanding that no sensitive information was accessed, that makes me no less concerned about the future risks that lie ahead.”
Alyssa Schwenk, a VDOE spokesperson, said the department has experienced “an influx in nonsensical or incomplete communication requests in the last several months, including to our FOIA portal.” She said in September, Anthropic “proactively flagged that three of these requests were from their AI agents. None of these attempts were fulfilled.”
Schwenk declined to publicly share the AI-generated requests.
“False requests have been easily identifiable due to generic names, fake addresses or email addresses, or nonsense contact information,” she wrote in the statement. “VDOE requests a photo ID for any non-verifiable FOIA submitter and will soon move to requiring ID for all submissions. At no point has any information been shared with a non-verified requestor … .”
In a follow-up email, Schwenk said the department will “continue to work with our colleagues and counsel to ensure that we are balancing our responsibility to protect personal data and information with our obligations under FOIA.”
VDOE houses a range of state information — but not student records. It does oversee the state’s Individualized Education Program, which “provides functionality as an IEP writer and compliance tool that affords educators the functionalities necessary to facilitate IEP meetings, develop IEPs, monitor student progress, and gather and submit data for state reporting purposes.”
Aaron Martin, an assistant professor of data science at the University of Virginia, said the lack of specificity in both Spanberger’s press release and VDOE’s response makes it difficult to fully understand the incident.
“What were the bots after? You know? Like, why on God's earth would these automated requests be submitting nonsense to the DOE FOIA portal,” he said.
Jennifer Guild, a spokesperson for the Virginia IT Agency, said the AI requests weren’t being considered an attempt to overwhelm the system — a tactic referred to as a “distributed denial of service attack.”
“If it is a nothing burger, by and large, then does it merit any response at all — besides maybe an internal re-evaluation of the identity checks, ensuring that the authentication requirements are fit for purpose,” Martin said. “You don't want to make the requirements too stringent, so that people don't submit information requests. But at the same time, if they're nonexistent, then there might be concerns around data being disclosed inappropriately. So, there is a balance there.”
Spanberger’s statement also references “federal inaction” on AI and references a letter the governor sent to federal leadership to register “deep concern regarding the unprecedented risks” of the technology.
Martin said he was encouraged by the move — that it signals the importance of considering AI’s impact on innumerable facets of society.
“I'm thinking about, for example, data centers. I'm thinking, for example, about companion chatbots and their impact on children,” he said. “Developing coherent policy for AI is hard because it really is impacting everything that we do.”