The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to analyze logs that include untrusted data sources such as user prompts, tool outputs, and terminal streams. This creates a surface where instructions embedded in the logs could potentially influence agent behavior during analysis.
Ingestion points: Data enters the agent's context through several files identified in SKILL.md, including events.jsonl and agenthost.log.
Boundary markers: The instructions provide general sensitivity warnings but lack explicit delimiters or specific directives for the agent to ignore instructions found within the log content.
Capability inventory: The agent has the capability to read local files, list directories, and execute the provided Python extraction script.
Sanitization: The process lacks content-level sanitization or escaping of the log data, which is standard for analysis tasks but remains a structural risk factor.
Command Execution: The skill involves the execution of a local Python script (extract.py) to prepare log bundles for analysis. This script is used to unpack archives into a temporary environment.
Defensive Extraction Logic: The included extract.py script implements robust security measures to mitigate common archive-based vulnerabilities. It includes checks to prevent ZipSlip (path traversal) and resource exhaustion attacks (Zip Bombs) by enforcing limits on entry counts, total extraction size, and individual file sizes. It also validates that paths are portable and do not use reserved names.

agent-host-logs