skill-library-sync
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides numerous shell commands for file management and system operations across sensitive directories like
~/.agents,~/.claude, and~/.hermes. - Evidence includes instructions to use
mv,ln -sfnfor symbolic links,rsync -a --deletefor mirroring, andtarfor archiving. - Mentions execution of a custom synchronization script
~/.agents/sync-to-claude.shwith flags that overwrite local files. - [EXTERNAL_DOWNLOADS]: Describes a process for downloading and integrating third-party code from external registries.
- Instructions detail using
npm packto download package payloads andtar xzfto extract them into the skill library for subsequent execution. - [INDIRECT_PROMPT_INJECTION]: The skill manages instruction sets (skills) from external sources, creating a potential surface for ingesting malicious commands into the agent environment.
- Ingestion points: Third-party skills downloaded via
npmor imported as "portable kits." - Boundary markers: The skill relies on manual review and documentation rather than programmatic delimiters to isolate external instructions.
- Capability inventory: The environment supports powerful file operations, symbolic linking, and script execution via
python3andnode. - Sanitization: Provides specific techniques for redacting sensitive information using
sedand verifying the absence of leaked tokens withrg(ripgrep). - Security Bypass: Instructs users or agents to add files to
~/.hermes/skillgate/self-owned.listspecifically to exempt them from automated security scanning by the platform's "watchdog" service. - [DATA_EXFILTRATION]: Explicitly identifies the risk of "fugas" (data leaks) when sharing skills with third parties.
- Acknowledges that managed files may contain sensitive identifiers such as client names, internal brands, tokens, and absolute file paths, and provides a procedure to identify and remove them.
Audit Metadata