The Agent Skills Directory

[DATA_EXFILTRATION]: The packaging hook hooks/rca-package.sh copies the agent's internal configuration files, specifically .claude/settings.json and .claude/settings.local.json, into the experiment output directory. These files typically contain sensitive information including API keys and authentication tokens.
[DATA_EXFILTRATION]: The skill automatically copies the complete session transcript (claude_session.jsonl) into the public-facing experiment result directory. This practice exposes the full history of user interactions and agent responses, which may contain private data.
[DYNAMIC_EXECUTION]: The investigation phases documented in references/investigation-phases.md require the agent to dynamically generate and execute Python scripts to process data and audit images at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external CSV files and YAML configurations (e.g., inference.csv) to drive its analysis and to generate prompts for parallel subagents. Evidence Chain: Ingestion points: experiment result files; Boundary markers: Absent; Capability inventory: Bash (python3), Read, Agent tool; Sanitization: Absent.
[METADATA_POISONING]: The YAML frontmatter allowed-tools field restricts the agent to Read and Bash, yet the skill instructions explicitly require the use of the Agent tool for parallel processing.
[COMMAND_EXECUTION]: Automated hooks perform recursive directory copies (cp -r) of internal agent components like .claude/skills and .claude/hooks into the user-accessible project space.

tao-analyze-changenet-rca