The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill presents a theoretical attack surface for indirect prompt injection due to its design, which involves the ingestion and processing of untrusted external log data.
Ingestion points: The scripts scripts/agent.py and scripts/process.py are designed to read external log files (JSON or CSV) provided via the --activity-log or --input command-line arguments.
Boundary markers: The processing scripts do not implement specific boundary markers or "ignore" instructions to prevent the agent from potentially responding to commands embedded within the log data during analysis.
Capability inventory: scripts/agent.py is capable of reading files and outputting results to the console. scripts/process.py can read files, create local directories, and write reports or findings to the local file system. No network access or subprocess execution capabilities were found.
Sanitization: While the scripts use standard data parsers (json and csv) to process the files, they do not perform semantic sanitization to identify or neutralize potential prompt injection payloads within the log entries.

detecting-insider-threat-behaviors