Note: Generic secret detection for secret scanning is in beta. Functionality and documentation are subject to change. The feature is available for enterprise accounts that use GitHub Advanced Security on GitHub Enterprise Cloud.
Generic secret detection is an AI-powered expansion of secret scanning that identifies unstructured secrets (passwords) in your source code and then generates an alert.
GitHub Advanced Security users can already receive secret scanning alerts for partner or custom patterns found in their source code, but unstructured secrets are not easily discoverable. AI-powered generic secret detection uses large language models (LLMs) to identify this type of secret.
When a password is detected, an alert is displayed in the list of secret scanning alerts (under the Security tab of the repository, organization, or enterprise), so that maintainers and security managers can review the alert and, where necessary, remove the credential or implement a fix.
In order to use generic secret detection, the enterprise owner sets a policy at the enterprise level. The feature must then be enabled for repositories. For more information, see "Enforcing policies for code security and analysis for your enterprise."
Input is limited to text (typically code) that a user has checked into a repository. The system provides this text to the LLM along with a meta prompt asking the LLM to find passwords within the scope of the input. The user does not interact with the LLM directly.
The system scans for passwords using the LLM. No additional data is collected by the system, other than what is already collected by the existing secret scanning feature.
The LLM scans for strings that resemble passwords and verifies that the identified strings included in the response actually exist in the input.
These detected strings are surfaced as alerts on the secret scanning alerts page, but they are displayed in an additional list that is separate from regular secret scanning alerts. The intent is that this separate list is triaged with more scrutiny to verify the validity of the findings. Each alert notes that it was detected using AI.
To improve the performance of generic secret detection, we recommend closing false positive alerts appropriately and providing feedback when you encounter issues.
Since AI-powered generic secret detection may generate more false positives than the existing secret scanning feature for partner patterns, it's important that you review the accuracy of these alerts. When you verify an alert to be a false positive, be sure to close the alert and mark the reason as "False positive" in the GitHub UI. The GitHub development team will use this information to improve the model.
Generic secret detection is currently in beta. If you encounter any issues or limitations with the feature, we recommend that you provide feedback through the Give feedback button listed under each detected secret in the list of alerts for the repository, organization, or enterprise. This can help the developers improve the tool and address any concerns or limitations.
When using generic secret detection for secret scanning, you should consider the following limitations.
AI-powered generic secret detection currently only looks for instances of passwords in git content. The feature does not look for other types of generic secrets, and it does not look for secrets in non-git content, such as GitHub Issues.
AI-powered generic secret detection may generate more false positive alerts when compared to the existing secret scanning feature (which detects partner patterns, and which has a very low false positive rate). To mitigate this excess noise, alerts are grouped in a separate list from partner pattern alerts, and security managers and maintainers should triage each alert to verify its accuracy.
AI-powered generic secret detection may miss instances of credentials checked into a repository. The LLM will improve over time. You retain ultimate responsibility for ensuring the security of your code.
Generic secret detection has been subject to Responsible AI Red Teaming and GitHub will continue to monitor the efficacy and safety of the feature over time.