vllm-pytorch-ci-triage
Pass
Audited by Gen Agent Trust Hub on Aug 19, 2026
Risk Level: SAFECREDENTIALS_UNSAFEPROMPT_INJECTION
Full Analysis
- Credential Handling Considerations: The skill instructs the user to store Buildkite and GitHub API tokens in specific files within the home directory (
~/.buildkite_token,~/.github_vllm_token). While the instructions emphasize secure file permissions (chmod 600), the agent's access to these files is a security consideration. - Indirect Prompt Injection Surface: The skill processes external data from CI logs and build reports to generate GitHub issue titles and descriptions. Because this data originates from external build processes, it represents a surface where unexpected content could influence the agent's output.
- Human-in-the-Loop Safeguards: The skill includes a mandatory step for the agent to draft issue content and wait for explicit user approval before posting to external APIs. This significantly mitigates the risk of automated actions driven by external data.
- Interaction with Well-Known Services: The skill interacts with official APIs for GitHub and Buildkite. These interactions are consistent with the skill's primary purpose of CI triage and issue management for the PyTorch and vLLM projects.
Audit Metadata