GitHub previews code scanning for AI applications and agent risks
GitHub announced code scanning for AI applications in public preview on October 28. The detector looks for AI-specific risks, including indirect prompt injection, sensitive information disclosure and excessive agency. A scanning feature is not a guarantee that an application is secure.
The announcement says scans run automatically when pull requests are created, alongside existing CodeQL analysis. The detector flags code paths where untrusted content can reach a model with access to tools, data or business systems. GitHub describes coverage as expanding, rather than claiming every possible agent vulnerability is detected.
Indirect prompt injection involves malicious instructions hidden in content an application reads, such as a file, webpage or issue comment. Sensitive information disclosure concerns missing authorization controls around private data. Excessive agency concerns permissions or capabilities broader than the task requires. These are categories of risk, not evidence that a particular repository has been compromised.
GitHub says the detector provides mitigation guidance in the pull request. Teams can inspect new alerts on the Conversation or Files changed tabs and review the affected code before merging. Guidance and a proposed change still require engineering judgment and testing.
The preview is available on github.com and requires a GitHub Advanced Security license. Both code scanning and scans for AI applications must be enabled in the repository. The announcement says the feature is not available on GitHub Enterprise Server at this time.
The source does not announce universal access for every GitHub account, a subscription price change or measured detection accuracy. Teams should check their license, repository settings and individual findings rather than treat preview availability as proof that scanning has already run or that every reported risk is exploitable.