suede-ship-copy
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection. Malicious instructions could be embedded in the external content the skill is designed to process, such as web pages fetched via URLs, repository files, or support transcripts.
- Ingestion points: The skill ingests untrusted data in
workflows/suede-ship-copy.jsduring theIntake(line 147),Research(line 237),Audit(line 332), andReview(line 538) phases. - Boundary markers: The workflow uses triple-quote (
""") delimiters to wrap external content within agent prompts (e.g., lines 228, 542, 706). It also includes negative constraints to sub-agents (e.g., "The audit never sees it", "not your business to check"), but these may not fully prevent adversarial instructions in the data from influencing the sub-agents. - Capability inventory: The agents are granted capabilities to read local files, fetch external URLs, and execute read-only
gitandghcommands (line 763). - Sanitization: The skill lacks explicit sanitization or escaping for the external data before it is interpolated into the JSON-formatted prompts sent to sub-agents.
- [COMMAND_EXECUTION]: The skill's workflow script (
workflows/suede-ship-copy.js) explicitly instructs sub-agents to execute system commands to gather data. - Intake Phase: Sub-agents are told to "curl the url, or read the file" to capture a baseline (line 173).
- Readiness Gate: Sub-agents are told they "may fetch urls, read files, and run read-only git and gh commands" (line 763).
- Context: These instructions are scoped to information retrieval (reconnaissance) and claim verification, which is consistent with the skill's stated purpose of fact-auditing copy.
- [DATA_EXPOSURE]: The skill is designed to read sensitive project information, such as
pricing config,changelogs, andtranscriptsto verify claims. While it includes boundaries to prevent publishing this data (line 823), the process involves aggregating various project-level data points into an "evidence record" and draft files.
Audit Metadata