Cybersecurity officials in California and Texas confirmed Thursday that they are among states participating in a new Anthropic cyber defense initiative designed to help state, local, tribal and territorial governments (SLTTs) use AI to identify system vulnerabilities before attackers can exploit them.
After months of working with and convening public-sector cybersecurity leaders, Anthropic has pledged up to $15 million in credits to help SLTTs use its Claude Security, Claude Code and the latest Opus 4.8 model to identify vulnerabilities and help remediate them, Danielle Cohen, Anthropic policy communications lead, state and local, said. The inaugural cohort launched Thursday and additional groups are expected to begin work over the summer.
California Chief Information Security Officer Vitaliy Panych said his agency, the California Department of Technology, was already using Claude and had very quickly discovered more than three dozen vulnerabilities.
“We are not waiting for the threat to mature before we act,” Panych said in an emailed statement. “Our teams are already using Claude across vulnerability management and security operations, and in early use one of our security engineers surfaced 40 vulnerabilities in a fraction of the time it would normally take.”
AI has significantly raised the stakes in cybersecurity, especially as frontier models from OpenAI, Google and Anthropic have been introduced. During a May congressional hearing, Center for Democracy and Technology Vice President of Policy Samir Jain warned that frontier AI models can uncover vulnerabilities that have been hidden for years, potentially allowing nation-state and criminal actors to identify weaknesses in government systems faster than agencies can address them.
An industry report in April said that AI assumptions about cybersecurity were already outdated, with the window between vulnerability discovery and exploitation shrinking to hours and, in some cases, minutes.
Anthropic’s partnership with SLTTs comes as experts have encouraged cyber leaders to use AI to find their own system weaknesses ahead of bad actors.
The effort is distinct from Project Glasswing, the restricted-access program built for critical infrastructure partners, Cohen said. Eligible entities, which also include special districts like port authorities or utilities and very large cities and counties, may apply for up to $100,000 in Claude credits with a six-month expiration. Anthropic’s applied AI team will provide support to selected public-sector teams through technical training and scheduled calls.
Texas Cyber Command, which is also part of the initial cohort, is using the opportunity to hone its responsiveness and flank emerging threats.
“There remains a lot of uncertainty and unknowns about what fully mature AI capability will mean when it comes to cybersecurity, but we do know that any number of bad actors will try to operationalize this, bringing a degree of disruption which will be difficult to deal with at speed,” its Cyber Chief Timothy “TJ” White said in a statement. “This teamwork … ensures we aren’t just reacting to potential disruption but building the agility needed to outpace emerging threats.”
Both Panych and White referenced the speed with which cyber criminals can act in the AI age. California, Panych said, “defends more people and more critical systems than almost any government in the country.”
“This partnership lets us meet that scale with speed, identify weaknesses before adversaries can exploit them, and set a standard other states can build on,” he said. “That is how you get ahead of these threats and stay ahead.”
Rae D. DeShong is a Texas-based staff writer for Government Technology and a former staff writer for Industry Insider — Texas. She has worked at The Dallas Morning News and as a community college administrator.