joseph gabriel ragoncin
news editor
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) uses Anthropic’s Mythos artificial intelligence model to scan federal government software for security vulnerabilities. Reuters reported that the deployment is part of a pilot program within the agency’s Attack Area Assessment Team.
The team uses Mythos to audit the source code of government systems, looking for flaws that could be exploited by cybercriminals and nation-state actors across a variety of applications.
People familiar with the effort told Reuters that Mythos had already identified multiple vulnerabilities during testing. The company did not specify the number or severity of the issues, or which government agencies or software stacks were affected.
The move puts Anthropic, which has come under separate scrutiny as a potential supply chain risk in the U.S. public sector, to the center of sensitive security workflows. It also highlights how quickly generative AI is moving from experimental use cases to high-stakes production environments within government.
Some experts say the choice of tool raises transparency issues. Some have focused on the broader impact AI-powered analytics could have on the long-standing security backlog in federal IT.
“The federal government seems undecided about what it thinks about AI in general and Mythos in particular. One week Anthropic is a supply chain risk, the next week CISA is giving Mythos the key to scanning federal code for vulnerabilities.” This inconsistency is bad enough, but it’s made worse by the fact that it’s unclear what Mythos is actually scanning: Is this code written by the government, or is it an internal bug? One, vendor bugs across federal systems are a supply chain issue, and the public has a right to know which one it is,” said Bronwen Aker, AI research and strategy analyst at Black Hills Information Security.
Security researchers note that large government environments contain extensive legacy code. Due to the size and age of the codebases involved, many have not undergone up-to-date systematic security reviews.
Jacob Krell, Suzuken’s senior director of secure AI solutions and cybersecurity, said the pilot only addresses part of the AI risk equation in software development.
“Using AI to scan legacy code for vulnerabilities and the AI generating vulnerable new code on the other end only solves half the problem. CISA pointing Mythos to the government codebase is a smart move. We’ve seen federal systems run code that hasn’t undergone serious security review for a decade, and models like Mythos can cover that volume in hours instead of months.
“The blind spot is on the production side. Every federal agency and contractor has developers who use AI assistants to write code, but those tools produce insecure output more often than secure output. Authentication flaws, hard-coded credentials, and lack of input validation all ship by default because the model optimizes for ‘will it run or not’ and skips the ‘is it safe?’
“When you combine both facts, you get a treadmill: Mythos discovers legacy bugs, teams apply patches, and AI coding tools introduce new vulnerabilities in the same repository at machine speed. The backlog does not shrink, it rejuvenates.
“Power grids and water systems are privately operated but subject to state oversight. CISA can’t just enforce federal regulations and say it’s done. If government agencies have scanning tools this powerful, the operators of critical infrastructure need access to them, too, because those are the systems that actually keep the lights on.
“We want CISA to combine this effort with the Secure Generation Standard for AI Coding Tools in Federal Development to extend scanning access to critical infrastructure providers, draining the pool while the hoses are still running,” said Jacob Krell, Senior Director of Secure AI Solutions & Cybersecurity at Suzu Labs.
While some see AI-based code scanning as an extension of long-standing safe development practices, it comes with new trade-offs in terms of complexity and noise.
“Software code review and analysis is not new. The reality is that most issues found are not exploitable unless very specific conditions are met. For example, the vulnerable function must actually be invoked and exposed to an attacker. “I think it’s likely that we’ll find many new issues that we weren’t able to identify. But the increased complexity may also limit exploitability. AI scans are likely to quickly produce a large amount of unworkable output, which experts will then have to review to identify the real risks.” Mr. Treynor, BHIS Penetration Tester and Antisyphon Instructor.
