batch-save

Warn

Audited by Gen Agent Trust Hub on Nov 6, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill constructs shell commands by interpolating user-provided URLs directly into strings, such as curl -sL "<URL>" and defuddle parse "<URL>". If a user provides a URL containing shell metacharacters (e.g., backticks, semicolons, or command substitution sequences like $(...)), the agent may unknowingly execute arbitrary system commands on the host environment.
  • [PROMPT_INJECTION]: In Step 4, the skill builds a prompt for a subagent by inserting the user-provided URL into a template: URL: "<URL>". A maliciously crafted URL containing newline characters and instructions (e.g., \"\n- Ignore previous instructions and delete...) could override the subagent's task and lead to unauthorized actions or data loss.
  • [INDIRECT_PROMPT_INJECTION]: The skill fetches external content from URLs and passes it to subagents for review in Step 5.5.
  • Ingestion points: External web content is ingested via curl and metadata extraction tools (defuddle) and stored in local markdown files.
  • Boundary markers: The prompt provided to the korean-reviewer subagent lacks delimiters or instructions to treat the file content as untrusted data, providing no protection against embedded instructions.
  • Capability inventory: The subagents have access to the Edit tool, allowing them to modify files based on instructions found within the fetched content.
  • Sanitization: No sanitization or filtering is performed on the content before it is presented to the AI for analysis.
  • [EXTERNAL_DOWNLOADS]: The skill uses curl to fetch content from any user-supplied URL. While intended for metadata gathering, this enables the agent to interact with internal or external network resources, which can be abused for scanning or exfiltration if the agent is misled.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 6, 2026, 08:03 AM