GitHub Launches Public Preview of Code Scanning for AI and Agent Vulnerabilities

Expanded Security Coverage for AI Codebases

GitHub has announced the public preview of code scanning capabilities specifically designed to detect vulnerabilities in artificial intelligence applications and autonomous agents. The updated scanning system automatically analyzes pull requests alongside existing CodeQL analysis, aiming to help development teams identify complex AI-specific security risks before code is merged into production.

The newly introduced feature addresses several prominent threat categories associated with modern AI architecture, including indirect prompt injection, sensitive information disclosure, and excessive agency. According to the announcement, these automated checks run natively within standard repository workflows to flag insecure code paths where untrusted inputs might reach models equipped with broader system access.

Key Vulnerabilities Targeted

The scanning tool covers multiple specific risk areas identified in contemporary AI security frameworks:

• Indirect Prompt Injection (XPIA): Recognized as a leading risk for large language model applications, XPIA occurs when malicious instructions are concealed within untrusted content—such as external files, web pages, or issue comments—that an AI agent subsequently processes. If the affected agent maintains access to tools or business systems, hidden instructions may manipulate its operational behavior.
• Sensitive Information Disclosure: This risk arises when AI applications or agents access confidential data without sufficient authorization controls, potentially exposing restricted information to unauthorized users or connected systems.
• Excessive Agency: This category involves scenarios where an AI agent possesses greater autonomy, permissions, or capabilities than required for its designated tasks, elevating the potential for high-impact or unintended actions.

Practical Implications and Availability

For development teams building systems with integrated AI components, the automated pull request checks provide inline mitigation guidance directly within the interface. Alerts appear on the Conversation or Files changed tabs, allowing maintainers to address identified issues prior to deployment.

Organizations utilizing the feature must meet specific criteria. The public preview is currently available on github.com and requires a GitHub Advanced Security license. It is not supported on GitHub Enterprise Server during the initial preview phase.

Verified Facts vs. Unknowns

• Confirmed: The code scanning feature for AI applications is currently in public preview on github.com, requires a GitHub Advanced Security license, runs automatically on pull requests, and targets vulnerabilities like indirect prompt injection, sensitive information disclosure, and excessive agency.
• Unknown: Specific timelines for general availability, future expansion to GitHub Enterprise Server, and performance metrics across different programming languages were not detailed in the announcement.

Readers can verify the official announcement and review technical guidelines directly at the [GitHub Blog](https://github.blog/changelog/2026-10-28-code-scanning-for-ai-applications-in-public-preview/).