The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection by processing external data from application logs.
Ingestion points: The scripts/agent.py utility ingests data from external logs via the --attack-log argument. Specifically, the analyze_attack_logs function parses JSON entries containing request_url and attack_params which are derived from untrusted user traffic.
Capability inventory: The script is designed for configuration and log auditing. It performs file system read operations on the config and log files and writes the final audit results to a JSON file. It does not perform network operations or execute shell commands.
Boundary markers: There are no explicit instructions or delimiters provided to the agent to treat data from the attack logs as untrusted or to ignore any embedded natural language instructions within those logs.
Sanitization: The script extracts and reports raw fields from the attack logs without performing sanitization or filtering to prevent potential prompt injection payloads from being processed by the agent when it reviews the generated report.

implementing-runtime-application-self-protection