review-changes
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of source code diffs and untracked files from the repository. This creates an attack surface where an attacker could embed malicious instructions (e.g., hidden inside code comments) to influence the behavior of the reviewer subagents.
- Ingestion points: Data enters the agent's context through
git diffoutput and the content of untracked files identified bygit ls-files. - Boundary markers: While the skill uses a "Reviewer brief" to define the scope and role of the subagent, it does not specify explicit structural delimiters (like XML tags) to separate the untrusted diff content from the system instructions.
- Capability inventory: The skill possesses the capability to execute shell commands (
git,npm,doistbot) and read the local filesystem. - Sanitization: The impact is mitigated by a mandatory human-in-the-loop step (Step 4: "Confirm each finding"), where the user must review and act on findings generated by the subagents, preventing autonomous execution of potentially malicious suggestions.
- [DATA_EXPOSURE]: The skill includes proactive security instructions in its "Prepare" phase. Step 1.2 explicitly mandates that the agent must check all files for credentials, tokens, and personal data before they are reached by reviewers or model providers. It also warns against using
git add -Ato prevent accidental staging of sensitive files like.env. - [EXTERNAL_DOWNLOADS]: The skill utilizes
npm ciwithin isolated worktrees to prepare the environment for running tests or code analysis. These dependencies are pulled from the standard Node package registry. The skill specifically references vendor-owned packages such as@todoist/cli-coreand@todoist/sdk. - [COMMAND_EXECUTION]: The skill executes standard development tools including
git,npm, and the vendor-specific CLI tooldoistbot. These commands are used for their intended purpose within the developer workflow and do not show signs of arbitrary or malicious execution.
Audit Metadata