The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/agent.py executes the aws CLI using subprocess.run to fetch CloudTrail events. It correctly uses argument lists rather than shell strings to mitigate injection risks.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external log data from CloudTrail which may contain attacker-controlled strings (e.g., User-Agents, resource names, or source IP strings).\n
Ingestion points: The query_cloudtrail_s3_events function in scripts/agent.py fetches data via the AWS CLI.\n
Boundary markers: No explicit delimiters or warnings to ignore instructions within the logs are present.\n
Capability inventory: The script performs command execution (AWS CLI) and file writing.\n
Sanitization: The script extracts specific fields using JSON parsing but does not sanitize content for potential downstream injection if the report is consumed by other tools.\n- [EXTERNAL_DOWNLOADS]: The skill instructions recommend the installation of standard Python packages boto3 and requests. Although imported, these packages are not actively used in the provided agent.py script, which relies on the AWS CLI instead.

analyzing-cloud-storage-access-patterns