oracle-cheatsheet
Fail
Audited by Gen Agent Trust Hub on Jul 29, 2026
Risk Level: HIGHCREDENTIALS_UNSAFECOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses sensitive conversation history logs located in
~/.claude/projects/. These JSONL files contain the full history of the assistant's interactions, which may include sensitive data, tool outputs, or credentials entered during the session. - [DYNAMIC_EXECUTION]: The skill executes a dynamically generated Python script using a shell heredoc (`python3
- <<'PYEOF'`) to parse session logs. While used for data processing, this pattern involves runtime script generation and execution.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from session history logs without robust sanitization.
- Ingestion points: Reads
~/.claude/projects/*.jsonlinSKILL.md. - Boundary markers: Absent; the extracted commands are passed directly to a subagent prompt.
- Capability inventory: Uses
bashfor file reading and writing the resulting cheat sheet. - Sanitization: Minimal; the Python script performs basic string splitting but does not filter for malicious instruction patterns within the history.
Recommendations
- AI detected serious security threats
Audit Metadata