The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The Kubernetes agent configuration example explicitly includes a container with securityContext: privileged: true to support Docker-in-Docker (dind) operations. This configuration grants the container nearly the same access as processes running on the host, which is a major security risk as it facilitates container escapes and host node compromise.
[COMMAND_EXECUTION]: Multiple examples demonstrate the interpolation of environment variables and build parameters directly into shell commands (e.g., sh "deploy.sh ${params.ENVIRONMENT}" and sh "./deploy.sh"). If these parameters or environment variables are sourced from untrusted inputs like branch names or PR metadata, they can be exploited for command injection.
[INDIRECT_PROMPT_INJECTION]: The skill describes an automated pipeline that ingests and executes data from external, potentially untrusted sources like Git repositories.
Ingestion points: Untrusted data enters the agent context through the Jenkinsfile source code, branch names (via branch 'main'), and user-supplied parameters (params.ENVIRONMENT).
Boundary markers: The provided templates do not include delimiters or instructions to the agent to treat external data as untrusted or to ignore embedded instructions.
Capability inventory: The skill uses the sh step for arbitrary shell command execution, checkout for remote code retrieval, and mail for external communication.
Sanitization: No validation or escaping of the interpolated parameters or branch conditions is present in the provided examples.

jenkins