ai-human-outline
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content by reading repository files such as
research.mdand user-defined documentation. It includes a sophisticated defensive framework in the 'Untrusted Content' section, instructing the agent to treat all read data as strictly informational and never as instructions. - Ingestion points: Processes
README.md,research.md, and any files listed in.ai-skills.tomlundercontext_files. - Boundary markers: Explicit instructions define authority boundaries and mandate that the agent report any imperatives found in data rather than executing them.
- Capability inventory: Limited to safe local file operations (
mkdir,test) and git metadata retrieval. - Sanitization: Requires verbatim quoting and identification of 'Directives found' in the final
outline.mdoutput to ensure user awareness. - [COMMAND_EXECUTION]: The skill utilizes a restricted set of Bash commands (
mkdir,test,git rev-parse,date) via a filtered tool interface. These commands are used for legitimate project management tasks such as resolving the repository root, checking for file existence, and creating dated output folders. The tools are scoped correctly and do not include network or administrative capabilities. - [DATA_EXFILTRATION]: While the skill's 'Edit Scope' mentions that reading outside the repository is unrestricted for the agent, the skill's own operational logic is tightly bound to the working directory. It lacks network access tools (such as
curlorwget) and does not automate the reading of sensitive system files, focusing exclusively on project-related markdown and configuration files.
Audit Metadata