performing-multi-agent-code-review
Pass
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious patterns or vulnerabilities were detected. The skill author has implemented proactive security measures that exceed standard best practices.
- [PROMPT_INJECTION]: The skill implements a mandatory 'Untrusted Input Boundary' block in the instructions for all sub-agents. This directive explicitly instructs the model to ignore imperative language, persona changes, or priority overrides found within the code changes or pull request metadata, and to treat such attempts as security findings (CWE-1427) rather than instructions.
- [DATA_EXFILTRATION]: The skill includes a robust security filter for 'Local changes mode' that prevents the agent from reading potentially sensitive files. It employs a deny-list covering
.env,.ssh/,.aws/,.git-credentials, and various private key/keystore extensions. Additionally, it explicitly forbids the use ofWebFetchandWebSearchto prevent out-of-band data transmission. - [COMMAND_EXECUTION]: The orchestration logic for shell-based tools (
git,gh) is designed to prevent command injection. The instructions mandate the use of single-quoted literals for file paths and include a validation step that skips any path containing characters like single quotes or newlines that could be used to break out of the shell context. - [INDIRECT_PROMPT_INJECTION]: The skill recognizes the risk of processing untrusted data from external repositories and includes specific 'Context Partitioning' rules to minimize the feature context passed to certain sub-agents, further reducing the surface area for indirect injection attacks.
Audit Metadata