The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes raw security telemetry and threat intelligence feeds which represent untrusted ingestion points controlled by external actors.
Ingestion points: The agent ingests data from Windows Event Logs (EventCodes 4625/4624), firewall/proxy logs, and threat intelligence lookups through the functions investigate_brute_force, correlate_across_sources, and check_threat_intel in scripts/agent.py.
Boundary markers: None. The skill lacks instructions or delimiters that would prevent an LLM from interpreting instructions embedded within the security logs as system commands.
Capability inventory: The skill allows for executing Splunk searches, retrieving sensitive telemetry, updating notable events via the Splunk ES API, and writing data to local storage (triage_report.json).
Sanitization: None. Log data and indicator descriptions are interpolated directly into SPL queries and passed to the agent context without escaping or validation.
[DYNAMIC_EXECUTION]: The skill constructs and executes Splunk Search Processing Language (SPL) queries at runtime based on user-supplied or data-driven inputs.
Evidence: In scripts/agent.py, functions like investigate_brute_force and enrich_with_asset_identity use f-strings to build executable SPL queries from the src_ip and username arguments. This is a low-risk pattern but establishes a surface for query-based injection if inputs are not strictly controlled.

triaging-security-alerts-in-splunk