problem-dive
Pass
Audited by Gen Agent Trust Hub on Sep 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and analyze untrusted external data, including system logs, trace/span chains, and user feedback reports, to perform root cause analysis. This creates a surface for indirect prompt injection if the analyzed logs or trace data contain malicious instructions intended to manipulate the agent's behavior during the investigation.
- Ingestion points: System logs (e.g.,
tail <log>), trace/span chains, and user feedback strings defined inSKILL.md. - Boundary markers: The instructions do not specify the use of delimiters or provide warnings to ignore embedded instructions within the logs being read.
- Capability inventory: The agent is expected to use tools for file reading, network searching for "due diligence" (尽调调研), and command execution for "experimental verification" (实测证伪).
- Sanitization: There are no instructions for sanitizing or escaping content from external logs or records before processing.
- [COMMAND_EXECUTION]: The methodology directs the agent to execute commands for evidence gathering (e.g.,
tail <log> | jq) and to perform "experimental verification" (实测证伪) to validate root cause hypotheses. This pattern involves executing shell commands or code based on data derived from the analysis process.
Audit Metadata