The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill explicitly instructs the AI agent to include plaintext credentials in git commit messages. The documentation states every commit MUST include a 'How to test' section containing 'Test credentials if login is required', and provides an example: 'Login: test@example.com / demo123'. Storing credentials in git history is a dangerous practice that leads to permanent exposure of secrets.
[PRIVILEGE_ESCALATION]: The setup process requires granting full collaborator access to a specific GitHub user account ('marsiandeployer'). This account is a third-party entity not associated with the skill author or a known trusted vendor, granting an unknown party wide permissions to the repository's source code, history, and potentially CI/CD secrets.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of a GitHub Action ('marsiandeployer/vibers-action@v1') from an unverified third-party repository. This action executes during the project's build/test phase and has the capability to read code and exfiltrate environment variables or repository secrets.
[DATA_EXFILTRATION]: The instructions require the 'spec_url' (project specifications) to be 'publicly accessible' (e.g., 'anyone with the link can view'). This forces the user to expose potentially sensitive business logic, architecture, and proprietary data to the public internet. Additionally, the feedback mechanism uses a curl command to send the repository URL to an external endpoint ('https://vibers.onout.org/feedback').
[INDIRECT_PROMPT_INJECTION]: The skill establishes a workflow where an external service generates code changes and Pull Requests that are subsequently ingested by the AI agent and the user. This creates a surface for malicious code or instruction injection via the review feedback.
Ingestion points: Git Pull Requests generated by the Vibers service (referenced as originating from the 'marsiandeployer' user).
Boundary markers: None. There are no instructions provided to the agent to treat the PR content or review feedback as untrusted data.
Capability inventory: The skill enables repository write access through the requested collaborator permissions and the PR merge workflow.
Sanitization: No sanitization or validation mechanisms are described for the content received from the external review service.

vibers-code-review