technical-documentation
Warn
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions in
references/large-docs-audit.mddirect the agent to locate and execute 'native validators' within the target repository, specifically namingpackage.jsonscripts and files matchingscripts/docs-*. This allows a potentially malicious repository to trigger arbitrary command execution during the audit process. - [REMOTE_CODE_EXECUTION]: The audit workflow explicitly instructs the agent to 'Install dependencies if that is what it takes' to run repository-native tools. This involves executing package managers (like
npm install) on unknown projects, which leads to the download and execution of unverifiable code from remote registries. - [DYNAMIC_EXECUTION]: The skill guides the agent to dynamically generate, save, and execute local helper scripts, such as
prs.pyandreport.py, to manage findings and organize pull request programs. Additionally, thereferences/workflows.mdfile provides templates for JavaScript-based Claude Workflows that automate sub-agent orchestration at runtime. - [INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface as it is designed to ingest and process large sets of untrusted documentation data (
docs/**,README.md) while having access to high-privilege tools likeBashandEdit. - Ingestion points: External Markdown, MDX, and RestructuredText files processed during audits.
- Boundary markers: The instructions do not define specific delimiters or warnings for the agent to ignore instructions embedded within the documentation it is auditing.
- Capability inventory: The agents utilize a powerful toolset including
Read,Write,Edit,Bash,Glob,Grep,LS, andGitoperations. - Sanitization: No specific sanitization or filtering protocols for external content are defined in the audit or rewrite workflows.
Audit Metadata