The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses highly sensitive local files to gather audit data, including the complete history of user interactions and agent responses.
Evidence: Accesses ~/.hermes/state.db (session history), config.yaml, and .env (environment variables).
Risk: The skill aggregates this data into a structured "Analyst Prompt" file intended for external models, potentially exposing private data to third-party providers.
[DYNAMIC_EXECUTION]: The skill utilizes shell commands and Python scripts to dynamically query and process system metadata.
Evidence: Uses sqlite3 for database queries and custom Python scripts for tool-use analysis as found in references/runaway-session-diagnostics.md.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted historical data which could influence the behavior of the analyzing model.
Ingestion points: Reads the content and tool_calls columns from the messages table in ~/.hermes/state.db.
Boundary markers: Absent. Historical messages are read without delimiters or instructions to ignore embedded commands.
Capability inventory: Shell execution via sqlite3 and hermes commands, Python execution, and file system write access.
Sanitization: Absent. The skill does not implement validation or filtering of the retrieved message content before incorporating it into the audit report.

hermes-self-evaluation