claude-md-manager
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where the agent reads and writes instructions to project-level metadata files (CLAUDE.md, AGENTS.md) and supplemental rule files (rules/*.md). Because these files are used to guide agent behavior, they represent a potential surface for indirect prompt injection if the files are modified by external, untrusted sources.
- Ingestion points: The skill targets the CLAUDE.md and AGENTS.md files, as well as the contents of the rules/ directory.
- Boundary markers: The instructions do not include specific delimiters or directives to ignore potentially malicious embedded instructions within these managed files.
- Capability inventory: The skill utilizes file system read and write operations to maintain the modular configuration files.
- Sanitization: No input validation or content filtering is specified for the data written to or read from the metadata files.
- [SAFE]: The skill includes informational links to official documentation from Anthropics (code.claude.com) and reputable developer blogs. These references provide best practices for memory management and are not associated with remote code execution or data exfiltration.
Audit Metadata