The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive internal agent files to calculate per-turn token overhead. Although no network exfiltration is present, the capability to read these files constitutes data exposure.
Evidence: scripts/audit_overheads.py and Python snippets in SKILL.md read from ~/.hermes/memories/USER.md, ~/.hermes/memories/MEMORY.md, and ~/.hermes/config.yaml to measure character and token counts.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from skill descriptions and memory files without implementing sanitization or boundary markers, creating a surface for potential instruction injection if these files contain malicious content.
Ingestion points: Skill frontmatter descriptions from all files in ~/.hermes/skills/, as well as content from MEMORY.md and USER.md processed by the audit scripts.
Boundary markers: None identified in the audit scripts or instructions to separate instructions from analyzed data.
Capability inventory: The skill has file system read access, SQLite database access, and the ability to write local files (e.g., generating PNG charts).
Sanitization: There is no evidence of filtering or escaping interpolated data read from the audited files.
[COMMAND_EXECUTION]: The audit procedure requires the execution of shell commands and local Python scripts to analyze the agent's state.
Evidence: SKILL.md provides multiple python3 -c one-liners and instructions for running the scripts/audit_overheads.py utility.

input-token-overheads