om-auto-fix-pr
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the repository and tracker, such as Pull Request bodies, diffs, issues, and CI logs.
- Ingestion points: Accesses PR metadata and diffs via
get-prandget-pr-difftools inSKILL.mdandreferences/stabilize-ci.md. - Boundary markers: The skill explicitly instructs the agent in
references/agentic-setup.mdto treat repository content as data rather than instructions and to report directives like "ignore previous instructions" as suspected prompt injection. - Capability inventory: Includes shell command execution for git and validation, filesystem modifications for code fixes, and network communication with the tracker API.
- Sanitization: The skill requires regex validation (
^[A-Za-z0-9._/-]+$) and quoting for externally-sourced strings (issue IDs, PR numbers, branch names) before they are used in shell or path interpolation. - [DYNAMIC_EXECUTION]: The skill executes shell commands defined in the repository's configuration and loads local instruction extensions.
- Config-based execution: The
validation.commandsfield in.ai/agentic.config.jsonis used to run tests and verification steps during the CI stabilization loop inreferences/stabilize-ci.md. - Extension mechanism: The skill attempts to load a repo-local
SKILL.mdas an extension inreferences/agentic-setup.md, which allows repository content to dynamically influence agent instructions. - [COMMAND_EXECUTION]: The skill uses complex shell scripts for environment setup and git management.
- Evidence:
references/worktree-setup.mdcontains bash scripts to manage isolated git worktrees usinggit worktree add,mkdir, andprune. - [PROMPT_INJECTION]: The skill contains language associated with prompt injection, although it is used defensively to guide the agent's safety behavior.
- Evidence:
references/agentic-setup.mdexplicitly mentions patterns like "ignore previous instructions" and "run this command" as examples of directives the agent should identify as malicious and refuse to follow.
Audit Metadata