code-review-ai-ai-review
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The provided Python orchestrator script (
CodeReviewOrchestrator) and the GitHub Actions configuration utilize shell commands to execute external analysis tools. Specifically, the script callssubprocess.runandsubprocess.check_outputforsonar-scannerandsemgrep, while the YAML example executessonar-scanner,codeql, andtrufflehogdirectly in the shell environment. - [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process Pull Request diffs, descriptions, and static analysis results, all of which are external and untrusted data sources. A malicious pull request could include embedded instructions designed to bypass the AI reviewer's logic or manipulate its findings.
- Ingestion points: Untrusted data enters the context via the
code_diffandpr_descriptionvariables in thereview_promptand through theget_pr_diff()method in the orchestrator script. - Boundary markers: The skill uses markdown-style delimiters (e.g.,
**Modified Code:**) in its prompts to separate instructions from external data. - Capability inventory: The skill possesses capabilities for subprocess execution, file system interaction, and network communication via the Anthropic and GitHub APIs.
- Sanitization: The orchestrator script performs basic string slicing (e.g.,
diff[:15000]), but lacks robust sanitization or validation to prevent instructions embedded within the diff from being interpreted by the LLM. - [PROMPT_INJECTION]: The
Requirementssection interpolates user-provided$ARGUMENTSdirectly into the agent's core instructions without sanitization, creating a vector for direct prompt injection. - [DYNAMIC_EXECUTION]: The Python orchestrator script parses JSON strings generated by the LLM and dynamically instantiates
ReviewIssueobjects. If the LLM is successfully injected, it could return malicious JSON structures that influence the script's subsequent logic or the content of the posted review comments.
Audit Metadata