CISA Reportedly Uses Anthropic’s Mythos to Audit Government Code Repositories

The U.S. Cybersecurity and Infrastructure Security Agency is reportedly using Anthropic’s Mythos AI model to scan government code repositories for security flaws. The effort points to a growing role for advanced AI in federal vulnerability discovery.

According to Reuters, CISA’s Attack Surface Evaluation team is using Mythos to review government software for bugs that could be exploited by foreign intelligence services or cybercriminal groups.

The report says the AI-assisted audits have already found a large number of vulnerabilities, but the affected systems, severity levels, and total amount of reviewed code remain undisclosed.

Why CISA’s reported Mythos use matters

CISA is responsible for helping defend federal civilian systems and reducing cyber risk across U.S. critical infrastructure. If the agency is using Mythos at scale, it could mark a major shift from manual code audits toward AI-assisted vulnerability discovery.

Traditional secure code review can take weeks or months on large repositories. AI models can analyze code faster, follow data flows across multiple files, and highlight suspicious logic that human teams can then verify.

However, AI scanning does not replace human review. It changes the workflow. Security teams still need to validate findings, prioritize remediation, avoid false positives, and prevent sensitive code or vulnerability details from being exposed outside approved environments.

Area	Reported detail	Why it matters
Agency	CISA	Federal cyber defense agency with government-wide security responsibilities
Team	Attack Surface Evaluation team	Conducts security assessments and hacking exercises across government systems
Model	Anthropic Mythos	Reportedly strong at vulnerability discovery and exploitation analysis
Target	Government code repositories	Could reveal exploitable bugs before attackers find them
Unknowns	Scope, affected systems, and severity	Limits public understanding of operational impact

Mythos has already drawn national security attention

Mythos has become one of the most closely watched AI cybersecurity systems in Washington. A separate Reuters report said the model identified vulnerabilities in highly sensitive U.S. government systems during classified testing.

That earlier report cited an Associated Press account of testing under Project Glasswing. Officials reportedly stressed that finding vulnerabilities quickly is not the same as proving the model could exploit every system on its own in the same timeframe.

The distinction is important. AI vulnerability discovery can be powerful, but government use still needs strong guardrails, controlled environments, audit logs, and human decision-making before fixes or offensive testing move forward.

How AI changes government vulnerability discovery

AI-assisted code review can help security teams inspect larger codebases and uncover patterns that older scanners may miss. These may include unsafe input handling, broken access control, injection risks, insecure deserialization, authentication logic mistakes, and risky dependency behavior.

The NIST Secure Software Development Framework recommends practices that reduce vulnerabilities in released software, mitigate the impact of unresolved flaws, and address root causes so issues do not recur.

AI tools can support those goals by finding potential defects earlier. Still, they should feed into a secure development process rather than operate as a stand-alone replacement for engineering, testing, and governance.

• AI can scan large repositories faster than manual review alone.
• Security engineers still need to confirm whether findings are real.
• Repository access must follow strict least-privilege rules.
• Vulnerability details should stay inside approved government environments.
• AI audit results need clear ownership, ticketing, and remediation deadlines.

What CISA may gain from Mythos

If used safely, Mythos could help CISA find vulnerabilities across federal codebases before attackers exploit them. That would support earlier remediation, stronger attack surface visibility, and better prioritization of high-risk software weaknesses.

The reported use also fits a broader government push toward proactive defense. Instead of waiting for external researchers, vendors, or adversaries to expose flaws, federal teams can search their own software more aggressively.

For agencies with large legacy systems, the potential value is clear. Many government applications rely on old code, complex permissions, and integrations across internal services. AI may help map those relationships faster than traditional tools.

Potential benefit	What it could improve
Faster review	Large repositories can be scanned more often
Better pattern detection	AI can flag logic flaws that rule-based scanners may miss
Earlier remediation	Teams can fix bugs before public disclosure or exploitation
Attack path analysis	Models may help connect small bugs into real exploit chains
Security workforce support	AI can help experts focus on the highest-risk findings

The policy question is as important as the technology

Mythos has also raised policy concerns because a model that helps defenders find vulnerabilities may also help attackers if access is poorly controlled. That dual-use problem sits at the center of the debate around advanced cybersecurity AI.

The NIST AI Risk Management Framework says AI risk management should consider risks to individuals, organizations, and society across the design, development, use, and evaluation of AI systems.

For government code audits, that means agencies need more than a powerful model. They need governance over who can run scans, what repositories the model can access, where outputs are stored, how findings are validated, and who can act on the results.

Export controls show how sensitive Mythos has become

Mythos and related Anthropic models have already triggered export-control concerns. The earlier Reuters coverage said the U.S. government ordered Anthropic to suspend exports of Mythos and Fable models worldwide and to foreign nationals, citing national security concerns.

Those restrictions reflected a difficult balance. U.S. agencies and trusted defenders may want access to the strongest AI security tools, while officials may fear that the same capability could accelerate vulnerability discovery for hostile states or criminal groups.

That debate is likely to continue as models become better at reading code, reproducing bugs, generating exploits, and recommending fixes. The challenge is to give defenders an advantage without creating a widely available offensive tool.

What federal agencies should watch

Agencies using AI for code audits need clear operating rules. Sensitive repositories, classified environments, and mission-critical software require strict isolation and logging.

The NIST SSDF provides a useful foundation because it treats secure software development as a lifecycle process. AI scanning should become part of that lifecycle, alongside threat modeling, secure coding, peer review, dependency control, and vulnerability response.

At the same time, the NIST AI RMF offers a governance lens for evaluating AI system risks. In this case, the model itself, the scanning pipeline, and the vulnerability output all need controls.

1. Limit model access to approved repositories and approved users.
2. Run scans in controlled environments with strong logging.
3. Keep sensitive code and findings inside authorized government systems.
4. Require human validation before findings are escalated as real vulnerabilities.
5. Track every finding through remediation, retesting, and closure.
6. Separate defensive scanning from any exploit development workflow.
7. Review model outputs for false positives, hallucinated bugs, and unsupported exploit claims.

Why this could become standard security practice

Federal agencies manage large software portfolios, and attackers increasingly look for weaknesses in public-facing systems, internal applications, cloud services, and software supply chains. AI could help defenders keep pace with that scale.

The reported CISA deployment may also influence private-sector security teams. If government assessors rely on AI for repository audits, large enterprises may adopt similar tools for internal red teams, application security reviews, and pre-release testing.

The key question is not whether AI can find bugs. It is whether organizations can safely handle the volume, sensitivity, and operational risk of the findings it produces.

Bottom line

CISA’s reported use of Mythos suggests that AI-assisted vulnerability discovery is moving from research and pilot programs into serious government security work. The model could help auditors find flaws faster across federal code repositories, but the public still lacks details about scale and impact.

The next phase will depend on governance. AI code audits need strict access control, secure handling of findings, human verification, and clear remediation workflows. Without those controls, the same tools that improve defense could create new risks around sensitive code and exploit knowledge.

FAQ