Pete Hegseth Called Anthropic a 'National Security Risk.' Now CISA Is Using It.

CISA is one of around 100 organizations approved to use Anthropic's Mythos model, which is far better than human penetration testers at detecting and exploiting software bugs.

The United States Cybersecurity and Infrastructure Security Agency (CISA) is now using artificial intelligence (AI) developer Anthropic's Mythos platform to "audit government software" -- a seeming rebuke of the Pentagon, which classified the company as a "national security risk" in February and attempted to institute controls on its use by government agencies.

CISA's Attack Surface Evaluation team has been employing the model to audit source code and identify flaws that could be exploited by cybercriminals or nation-state actors, Reuters first reported last week.

The initiative is apparently part of a pilot program to evaluate whether AI can accelerate software security reviews across government systems. According to the Reuters report, Mythos successfully identified multiple vulnerabilities during testing, although specifics on the number of vulnerabilities, severity, or affected software were not disclosed.

What to Know About Claude Mythos

The Claude Mythos platform is an extremely advanced large language model (LLM) that hasn't been released to the general public due to concerns that its software engineering capabilities could be too dangerous. The software has already been used to identify thousands of zero-day vulnerabilities in every major operating system and web browser, including decades-old bugs.

As the AI can spot even the most subtle bugs, it has the capability to self-correct any coding mistakes, giving it clear commercial applications. The main risk is that in the wrong hands, it could also autonomously execute exploits at scale and speed that far exceed any human capabilities.

"Software code review and analysis are nothing new. Realistically, most issues found are not exploitable without very specific conditions being met (i.e., the vulnerable function needs to actually be invoked and exposed to the attacker in order to be abused)," explained Chris Traynor, penetration tester at Black Hills Information Security (BHIS) and instructor at Antisyphon.

Traynor told The National Interest via email that he believes that AI vulnerability scanning will likely find many new and novel issues that were simply too complex to identify with legacy tools before.

"But added complexity can cause limitations in exploitability," Traynor said. "AI scanning will likely produce a lot of unactionable output very quickly that will need to be reviewed by experts to find the real risks."

AI Is Far Better Than Humans at Finding Bugs -- for Better and Worse

Software bugs are virtually inevitable and incredibly common in development. Complex systems require thousands of lines of code, and anything from minor typos to miscommunications to logical missteps can lead to bugs. This introduces unpredictable variables.

Code also doesn't exist in a bubble, nor is it created in one. Changes in operating systems, web browsers, and third-party services can break functioning software, often without warning.

More worrisome is that such software bugs serve as the foundation for almost all exploits in cybersecurity. A bug has the potential to create unintended weaknesses, such as how a program handles data or permissions. Many software vulnerabilities can go unnoticed, often hidden for years.

The danger is that AI can find those quickly. It is an extremely useful, but essentially neutral, tool: a developer or "white hat" penetration tester can use AI to help patch faulty code, or a "black hat" hacker can use it for nefarious purposes.

"AI finding vulnerabilities in federal code at scale is interesting, but the harder question is what happens after the finding. A vulnerability that exists in a library no one calls, behind a network segment no one reaches, is not the same problem as one sitting in a critical authentication path," said Seemant Sehgal, founder and CEO of cybersecurity provider BreachLock.

"Without validating exploitability and reachability, every finding lands with the same weight, and that creates its own kind of risk," Sehgal told The National Interest. "The real test of this program is whether the output helps prioritize action or just expands the backlog."

The US Government Has a Mercurial Approach to AI

The federal government is already using AI extensively, deploying more than 3,600 distinct AI programs across its agencies. But the Trump administration is less clear about how it believes AI should be used in security matters.

"The federal government can't seem to decide what it thinks about AI in general, or Mythos in particular," Bronwen Aker, AI research and strategy analyst at BHIS, told The National Interest. "One week Anthropic is a supply-chain risk, the next week CISA is handing Mythos the keys to scan federal code for vulnerabilities."

The intra-government squabbles over Anthropic are by now a matter of public record. In February, the Pentagon called for Anthropic to be banned across the entirety of the federal government. Four months later, in June, the Department of Commerce instituted an export ban on two powerful Anthropic models, Fable and Mythos, leading Anthropic to shut both of them down.

Earlier this month, the Department of Commerce lifted the export ban, and Fable has been made available to the general public once more. Mythos had been strictly limited to just around 100 vetted US organizations focused on critical infrastructure and cybersecurity. That seems to include CISA, which is employing the platform in the software audit.

"That inconsistency would be bad enough to start with, but because it's not clear what Mythos is actually scanning, it's much, much worse," Aker told The National Interest in an email.

It remains unclear if it is government-written code or software built by third-party contractors and vendors.

"In-house bugs are one problem," warned Aker. "Vendor bugs running across federal systems are a supply chain problem, and the public has a right to know which one this is."

AI-Generated Code Is Part of the Security Problem

Part of the problem is that at the same time AI scans for vulnerabilities in legacy code, it also generates new code on the other end -- code with many of the same flaws in it as human-generated code. In other words, using AI to patch bugs is only solving half the problem.

Jacob Krell, senior director for secure AI solutions and cybersecurity at security provider Suzu Labs, observed that CISA pointing Mythos at government codebases should be seen as a smart move, but perhaps only the opening move in what could be a long game.

"I've seen federal systems running code that hasn't had a serious security review in a decade, and a model like Mythos can cover that volume in hours instead of months," Krell told The National Interest. "Every federal agency and contractor also has developers writing code with AI assistants, and those tools produce insecure output more often than secure output. Authorization flaws, hardcoded credentials, missing input validation, all shipping by default because the models optimize for 'does it run' and skip 'is it safe.'"

The danger is that when combined, the result is akin to a treadmill that isn't so easy to get off of. In this case, Mythos finds legacy bugs, teams patch them, but then AI coding tools introduce fresh vulnerabilities even as they fix the old ones.

That puts everything from power grids to water systems, which may be privately run, still squarely in the crosshairs of hackers.

Krell suggested that CISA can't simply harden federal code and call it done.

"If the agency has a scanning tool this capable, the operators running critical infrastructure need access to it too, because those are the systems that actually keep the lights on," he said. "I'd want CISA to pair this initiative with secure-generation standards for AI coding tools in federal development, and extend scanning access to critical infrastructure operators. We are draining the pool while the hose is still running."

About the Author: Peter Suciu

Peter Suciu has contributed to dozens of newspapers, magazines and websites over a 30-year career in journalism. He regularly writes about military hardware, firearms history, cybersecurity, politics, and international affairs. Peter is also a contributing writer for Forbes and Clearance Jobs. He is based in Michigan. You can follow him on Twitter:@PeterSuciu. You can email the author: [email protected].