workflow-big-feature
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill includes instructions intended to override the agent's default processing behavior, such as 'read that file in full yourself (every line, with no line limit)' and instructions to disregard host refusal of tool calls in favor of direct file reading.
- [COMMAND_EXECUTION]: The skill documentation references the execution of a local maintenance script using Node.js:
node .claude/scripts/lib/workflow-skill-contract.cjs --write. - [INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect instructions through its integration of untrusted data sources and high-privilege capabilities.
- Ingestion points: The skill processes results from external web research via the
$web-researchstep, user-provided prompts, and various project-level configuration files such asworkflows.jsonandproject-config.json. - Boundary markers: There are no explicit delimiters or 'ignore embedded instructions' warnings provided for the data retrieved from external web research or project documentation.
- Capability inventory: The skill possesses the authority to spawn sub-agents (
spawn_agent), execute shell commands (e.g.,node), and perform persistent file system writes for run reports and todo lists intmp/reports/. - Sanitization: The instructions do not mention any sanitization, escaping, or validation protocols for data ingested from remote research sources before it is processed by the agent or sub-agents.
Audit Metadata