States brace for AI-driven cyber attacks

Author:

Using an unordered list, summarize the following article with between 4 and 8 key points. Rapid advancements in artificial intelligence are pressing state and local governments to shore up their cybersecurity operations to protect against hacks.

Their concern is that frontier AI models in the hands of bad actors could be used to unleash sophisticated attacks that exploit previously unknown vulnerabilities.

“The world is dangerous now, and it’s about to get way more dangerous as these frontier models become more widely available,” Alan Fuller, Utah’s chief information officer, told Pluribus News.

A recent survey of state chief information security officers highlighted the peril: Only 22% said they felt confident in their ability to protect public data, down from 48% in 2022.

Human hackers used to have to hunt for gaps in cybersecurity walls. AI now makes it possible to deploy an army of virtual agents to do the work, exponentially raising the chances of surprise intrusions known as zero day attacks.

The danger was highlighted in September when a suspected Chinese state-sponsored group used Anthropic’s Claude Code tool to target tech companies, banks, chemical manufacturers and governments.

Anthropic called it “the first documented case of a large-scale cyberattack executed without substantial human intervention.” It issued a report that warned “the barriers to performing sophisticated cyberattacks have dropped substantially.”

“Threat actors can now use agentic AI systems to do the work of entire teams of experienced hackers with the right set up, analyzing target systems, producing exploit code, and scanning vast datasets of stolen information more efficiently than any human operator,” the report said.

The issue was amplified in April when Anthropic said its new AI model, Claude Mythos, was too dangerous to release publicly because of the implications for cybersecurity. Instead, the company shared Mythos with more than three dozen tech companies to give them a “head start” on finding and fixing security gaps, the New York Times reported.

Project Glasswing is Anthropic’s initiative to “secure the world’s most critical software” after Mythos found “thousands of high-severity vulnerabilities,” including across major operating systems and web browsers. The project has so far found more than 10,000 serious security gaps and was expanded last week to include infrastructure operators in more than 15 countries.

The warnings about Claude Mythos prompted President Donald Trump last week to issue an executive order on AI and security. It includes a requirement that the Homeland Security secretary “facilitate access to cybersecurity tools and services” for state and local governments and operators of critical infrastructure such as utilities and hospitals.

OpenAI announced last month it was offering a “limited preview” of GPT-5.5-Cyber to entities responsible for the cybersecurity of critical infrastructure.

The immediate priority has been to get frontier tools into the hands of enterprise software developers and critical infrastructure operators to help them find gaps in their security before bad actors do.

State and local governments will benefit from those efforts. But the public sector has its own systems potentially vulnerable to hackers, including nation-state actors hostile to the United States and criminal syndicates.

Fuller said potential intruders scan Utah’s computer networks more than a billion times a day.

The state deploys an AI-powered tool to scan its network traffic and server logs daily, looking for suspicious traffic and malicious code. But Fuller worries it isn’t enough, especially if experts are correct that states have six to nine months before hackers gain access to frontier models.

“What’s happening is an arm’s race,” said Fuller, who serves as vice president of the National Association of State Chief Information Officers’s executive committee. “Can we build up our defenses fast enough to stay ahead of the attacks that we’re pretty sure are coming with frontier AI models.”

An Anthropic spokesperson told Pluribus News the company has worked with state, local, tribal and territorial governments, including holding a series of cybersecurity briefings in May that drew more than 100 top public sector tech officials.

The company plans to announce this week a new cyber defense program tailored specifically to the public sector, using generally available Claude models to identify and fix vulnerabilities.

The recent survey conducted by NASCIO in partnership with Deloitte painted a concerning picture: State cybersecurity budgets aren’t keeping up; confidence in the ability to ward off attacks is waning; and states’ reliance on third-party vendors heightens vulnerabilities.

“State chief information security officers are protecting public data systems at a time when cyber threats are growing in sophistication, as foreign adversaries, sophisticated hackers and cybercriminals are increasingly using new artificial intelligence-based tools to probe for weaknesses,” the report said.

AI-enabled attacks are a top-three concern for state information security officers, along with phishing attempts and security breaches involving third-party contractors.

State cybersecurity officers are also worried about the risk to local governments and public higher education institutions, which sometimes have access to state systems. The survey found 26% confidence that state information assets are protected from cyberthreats originating within local government and public higher education, down from 59% in 2022.

Despite the bleak outlook, chief information security officers in 23 states reported in the survey that they are already using generative AI to shore up their cybersecurity operations, with 21 more states planning to do so in the next year.

Additionally, approximately one-fifth states are adopting a “whole-of-state” approach to cybersecurity, which involves the state providing umbrella support to entities outside of state government.

But funding remains a limiting factor, as states face budget shortfalls that make new investments in security a tougher sell in state legislatures. Cybersecurity budgets are largely flat or falling amid state spending shortfalls and federal funding expiring, the survey found.

Last month, state leaders urged Congress to reauthorize the $1 billion State and Local Cybersecurity Grant Program.

“We are under-resourced, and that’s part of what is causing us fear,” Fuller said.

Another potential line of defense are state laws requiring frontier AI developers to mitigate the risk of catastrophic harms, including widescale cybersecurity attacks. California and New York already have first-in-the-nation laws on the books, and a recently passed bill in Illinois is headed to the governor.

Read more: Blue states push toward de facto national AI safety standard

Tags:
President Donald Trump, Anthropic, Claude Mythos, OpenAI, Cybersecurity operations, Frontier AI models, Project Glasswing, AI-enabled attacks, artificial intelligence, Claude Code, Alan Fuller, state and local governments, Utah, GPT-5.5-Cyber