Key Points:
- The U.S. Cybersecurity and Infrastructure Security Agency is using Anthropic’s unreleased AI model, Mythos, to audit government software.
- CISA’s Attack Surface Evaluation team is employing the powerful AI to scan code repositories for vulnerabilities exploitable by foreign spies.
- Unveiled in April 2026, the highly advanced Mythos model can autonomously discover and exploit zero-day bugs in operating systems.
- The deployment comes as Anthropic navigates strict federal regulations, including a recent order to suspend certain model access for foreign nationals.
The leading United States cyber defense agency has quietly integrated one of the world’s most powerful and controversial artificial intelligence systems into its defensive operations. The Cybersecurity and Infrastructure Security Agency has gained full access to Anthropic’s unreleased AI model, Claude Mythos, using it to scan and audit critical government software repositories. The move signals a major shift in how federal authorities defend public infrastructure from highly sophisticated foreign espionage and cybercriminal networks. It also marks a significant victory for the cybersecurity agency, which had struggled for months to secure access to the restricted technology.
The highly sensitive scanning operations are being conducted by the agency’s internal Attack Surface Evaluation team. This specialized unit acts as an elite digital defense group, carrying out proactive security assessments, simulated hacking exercises, and code reviews across multiple federal departments. By deploying the advanced artificial intelligence model to scan public-facing codebases and internal software libraries, the team can analyze complex systems in a fraction of the time required by human engineers. The automated audits aim to identify latent security flaws, outdated configurations, and exploitable bugs before hostile state-sponsored hackers can weaponize them.
The decision to use this specific model is highly significant because of the system’s unprecedented offensive capabilities. When developer Anthropic first unveiled the Claude Mythos Preview model in April 2026, it stunned the technology sector by demonstrating an ability to autonomously discover and exploit zero-day vulnerabilities in every major operating system and web browser. Recognizing that the software could easily upset the global cybersecurity balance of power if leaked, the developer chose to hold the model back from public release. Instead, the firm kept the code behind closed doors, restricting access to a highly vetted group of defensive partners.
To put these autonomous capabilities to work safely, the artificial intelligence startup established a defensive early-access coalition called Project Glasswing. Backed by up to $100 million in computing credits, the program partners with approximately 50 critical infrastructure organizations and software vendors, including technology giants like Apple, Microsoft, Google, Amazon Web Services, and Nvidia. Under the program’s strict guidelines, partners use the model to scan their own proprietary codebases and coordinate patches. This defensive model establishes a strict 90-day disclosure clock, allowing software developers a crucial head start to patch severe vulnerabilities before the details become public.
The agency’s current deployment represents a major turnaround from its standing earlier this spring. Shortly after the model’s debut in April, the national cyber defense agency was on the outside looking in, lacking access to the tool even as other federal bodies like the Department of Defense and parts of the intelligence community began active testing. This initial exclusion drew sharp criticism, with many arguing that leaving the country’s primary civilian cyber defense agency out of the loop was a major strategic oversight. The successful integration of the tool by the Attack Surface Evaluation team resolves this bureaucratic bottleneck, aligning the agency with the nation’s top defensive efforts.
The technical capabilities driving this government enthusiasm are backed by rigorous independent testing. In evaluations conducted by the United Kingdom AI Security Institute, the unreleased model successfully solved 73% of expert-level capture-the-flag cybersecurity tasks. This performance vastly outpaced previous general-purpose language models, which typically struggle to handle complex, multi-step hacking scenarios that require deep logical reasoning and recursive self-correction. By automating the discovery of complex zero-day bugs, the tool compresses the diagnostic window from weeks of manual auditing to just a few hours.
The real-world impact of the defensive scanning program has already yielded massive results across the tech ecosystem. Since the launch of Project Glasswing, the defensive coalition has identified more than 10,000 previously unknown, high-severity vulnerabilities across major software applications, operating systems, and open-source libraries. While finding these flaws represents a major success for defensive teams, it also highlights the immense scale of latent risk currently embedded in global digital infrastructure. As similar autonomous capabilities eventually proliferate to open-source models, organizations must accelerate their patching tempos to prevent widespread exploitation.
The agency’s adoption of the tool occurs against a highly complex regulatory backdrop, characterized by an ongoing standoff between the artificial intelligence developer and the White House. Federal authorities have recently tightened export controls and security requirements on advanced frontier models, ordering the developer to suspend access to its newest high-capability models, including Fable 5 and Mythos 5, for foreign nationals. While these restrictions have created administrative hurdles for the startup’s international partnerships, the company has worked closely with federal regulators to secure the necessary clearances to reactivate its models for authorized government defense programs.
Ultimately, the integration of autonomous vulnerability discovery into federal cyber defense marks the beginning of a new era in national security. As the velocity of digital threats continues to accelerate, relying solely on human software engineers to identify and patch flaws is no longer a viable strategy. The coming months will reveal whether the cybersecurity agency can successfully leverage these advanced diagnostic capabilities to build permanent, durable cyber resilience across the federal government. For now, the deployment of this unreleased technology proves that the battle between digital defenders and attackers will increasingly be fought using the most advanced models available.
