trace-grader
Warn
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill accesses sensitive information by reading session transcripts and audit logs located in directories such as
~/.claude/projects/and.claude/audit/. These files contain a comprehensive record of agent actions, which may include environment variables, secrets, or sensitive code blocks accessed or displayed during the agent's session. - [INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection by processing external, untrusted data from session traces.
- Ingestion points: Processes
traces.json, audit logs, and session transcript files via the--transcriptflag. - Boundary markers: The skill instructions do not define delimiters or specific warnings to ignore instructions embedded within the trace data being graded.
- Capability inventory: The skill executes a local Python engine
harness/scripts/trace-grader.pyto process the data. - Sanitization: There is no documented validation or sanitization of the contents within the traces or transcripts before they are parsed by the grading script.
Audit Metadata