bmad-advanced-elicitation
Warn
Audited by Gen Agent Trust Hub on Nov 12, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill is designed to execute "Literal instructions" found in the configuration fields
activation_steps_prepend,activation_steps_append, andon_complete. These instructions are run via the shell during different stages of the skill's lifecycle. - [DYNAMIC_EXECUTION]: The skill frequently uses
uv runto execute local Python scripts (resolve_customization.py,resolve_config.py,pick_methods.py,roster.py). The parameters for these executions are dynamically resolved from configuration files and piped between processes, which is a powerful execution primitive. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (the "target") and provides it to subagents for analysis, creating a potential vulnerability surface.
- Ingestion points: Untrusted data enters via user-pointed files, chat history staged to
target.md, and files loaded throughpersistent_factsglobs. - Boundary markers: The skill attempts to mitigate injection by using a structured "Brief" format that constrains the subagent's context and provides explicit instructions to "return only that as text" and "invoke no skill."
- Capability inventory: The skill possesses significant capabilities including file read/write, subagent management, and shell command execution via
uv run. - Sanitization: No explicit sanitization or escaping of the target content is performed before it is interpolated into the subagent briefs.
- [DATA_EXFILTRATION]: The
persistent_factsconfiguration field allows for the automatic loading of files or globs from the project directory. If an attacker influences this configuration, it could lead to the exposure of sensitive files to the AI's context.
Audit Metadata