The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data from multiple sources, including canonical traces.json files, audit logs (.jsonl), and Claude Code session transcripts. These data sources represent the primary attack surface where malicious instructions could be embedded within tool observations or step descriptions. The skill does not explicitly define boundary markers (e.g., delimiters) or specific sanitization routines for the content of these traces. However, the skill's capabilities are primarily limited to local processing, rule-based checking, and reporting via stdout and exit codes, which significantly limits the potential impact of such an injection.

trace-grader