dig

Pass

Audited by Gen Agent Trust Hub on Jul 29, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It reads historical session data from .jsonl files in ~/.claude/projects/ and renders session summaries directly into the agent's context. If a previous session contains malicious instructions (e.g., in a message or a summary field), the agent may inadvertently follow them when processing the output of the /dig command.\n
  • Ingestion points: scripts/dig.py parses all .jsonl files within the configured project directories.\n
  • Boundary markers: The output is formatted as a Markdown table without explicit delimiters or warnings to the agent to ignore embedded instructions within the session summaries.\n
  • Capability inventory: The skill possesses the ability to execute shell commands and read local files.\n
  • Sanitization: The script truncates summaries to 80 characters, which provides minimal protection against payload execution but does not neutralize the threat.\n- [COMMAND_EXECUTION]: The skill performs shell command execution to determine the environment state and repository metadata.\n
  • Evidence: SKILL.md uses ls, sed, and date to resolve project paths and timestamps. scripts/dig.py uses subprocess.run to call date for timezone detection and ghq for repository name resolution. These are standard system utilities used for their intended purposes.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 29, 2026, 02:56 AM
Security Audit — agent-trust-hub — dig