The Agent Skills Directory

[PROMPT_INJECTION]: No evidence of prompt injection or behavior override patterns was detected. The instructional language is professional and strictly adheres to the stated defensive DevSecOps purpose.
[DATA_EXFILTRATION]: No evidence of data exposure or exfiltration. The skill provides clear instructions on protecting credentials, such as advocating for OIDC federation over stored cloud keys and encrypting infrastructure-as-code state files.
[EXTERNAL_DOWNLOADS]: The skill references and provides instructions for using various external security tools and repositories. These are well-known, trusted industry-standard projects, including Sigstore, Kyverno, Anchore Syft/Grype, and Aqua Security Trivy. These references are documented neutrally as they are central to the skill's defensive objectives.
[COMMAND_EXECUTION]: The skill includes shell command examples and Dockerfile snippets for building and auditing pipelines. These are provided as educational best practices and audit checklists for the agent to recommend to users, rather than malicious execution routines.
[INDIRECT_PROMPT_INJECTION]: The skill involves processing external configuration files like CI/CD workflows and Dockerfiles, which represents an inherent attack surface.
Ingestion points: Analyzes user-provided files such as .github/workflows/*.yml, Dockerfile, and package manifests during BUILD and AUDIT modes.
Boundary markers: Explicitly instructs the use of environment variable indirection (Rule 1.5) to separate untrusted data from execution contexts.
Capability inventory: Provides instructions for generating build artifacts and modifying deployment configurations.
Sanitization: Recommends linting tools (zizmor, actionlint) and shell-safe practices to mitigate injection risks from untrusted PR metadata.

sota-devsecops