3 Sources: CISA Deploys Anthropic Mythos AI to Audit Government Code for Bugs

Federal defenders are turning to artificial intelligence to stay ahead of adversaries, and a new exclusive reveals that CISA is now using Anthropic's Mythos AI to proactively audit government software for vulnerabilities that could be exploited by criminals or nation-state spies. The deployment by the Attack Surface Evaluation team puts advanced language models directly into the heart of public-sector cybersecurity, scaling code analysis far beyond what human teams alone can achieve.

The U.S. Cybersecurity and Infrastructure Security Agency has quietly started using Anthropic’s AI model Mythos to audit federal code repositories for vulnerabilities, according to three sources familiar with the matter. This exclusive report reveals a deepening, though still uneasy, embrace of the AI startup’s tools by the very government that only months ago sought to blacklist it. The development underscores both the escalating demands of digital defense and the high-stakes politics surrounding advanced AI. CISA’s Attack Surface Evaluation team, the internal group responsible for probing federal networks for weaknesses, is employing Mythos to automate the tedious but critical task of scanning millions of lines of code for bugs that foreign spies or cybercriminals could exploit. While the scale and severity of vulnerabilities uncovered remain cloaked in operational secrecy, two sources indicated that the audits have already flagged a large number of issues—an outcome that, if confirmed, would represent a significant validation of AI’s role in proactive cyber defense.

“

Cybersecurity and Infrastructure Security Agency has quietly started using Anthropic’s AI model Mythos to audit federal code repositories for vulnerabilities, according to three sources familiar with the matter.

The backdrop is anything but ordinary. Anthropic, a San Francisco-based AI company founded by former OpenAI researchers, has long positioned itself as a safety-first developer, embedding constitutional AI principles into its models to refuse harmful outputs. That ethos collided with the Pentagon in early 2026, when the company refused to remove safeguards that prevented its AI from being used for autonomous weapons or domestic surveillance. The Department of Defense responded in February by slapping Anthropic with a formal supply-chain risk designation—a label previously reserved for foreign firms suspected of facilitating espionage. The move effectively barred the company from federal contracts and cast a shadow over its enterprise ambitions. That extraordinary action was blocked by a federal judge in March, a judicial stay that allowed Anthropic to continue operating while the legal and political dust settled. By July, the relationship had thawed sufficiently for CISA to quietly deploy Mythos on sensitive government systems, an irony not lost on observers: the agency entrusted with defending the nation’s cyber infrastructure is now relying on the same AI maker once deemed a supply-chain threat.

The implications for federal cybersecurity are substantial. Manual code review, the traditional method for vetting software used by agencies, is slow, expensive, and unable to keep pace with the volume of new patches and applications. AI models like Mythos can process entire repositories at machine speed, flagging suspicious patterns, outdated libraries, or hardcoded credentials that might otherwise slip through. By integrating AI into its attack surface assessments, CISA potentially multiplies its capacity to harden government systems against adversaries ranging from state-sponsored hacking groups to ransomware gangs. Yet the move also introduces new risks. Large language models are prone to hallucinations—generating plausible but incorrect findings—and can miss context-specific vulnerabilities that a human auditor would catch. Moreover, the opaque nature of AI decision-making raises questions about accountability when a missed bug leads to a breach. The lack of public detail about how many bugs were found, their severity, and how the AI’s output is validated highlights the nascent and still experimental character of this deployment.