The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from log files and database tool outputs, creating a surface for indirect prompt injection when the resulting reports are read by an agent.
Ingestion points: scripts/analyze_logs.py and scripts/hermes_log_window.py ingest raw log files, while scripts/state_failures.py reads tool call results and messages from the state.db SQLite database.
Boundary markers: The generated Markdown and JSON reports do not use explicit delimiters or specific instructions to the agent to treat summarized log content as untrusted data.
Capability inventory: The skill's scripts are restricted to local file system access for reading logs and writing reports; no network exfiltration or dynamic code execution capabilities were detected in the analysis of the skill's logic.
Sanitization: scripts/analyze_logs.py implements a normalize_message function that replaces URLs, IP addresses, and numbers with placeholders, which partially mitigates the exposure of sensitive strings, although full log snippets are still included in crash and failure reports.

log-analyzer