create-write-like-me
Warn
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The instructions in
SKILL.md(Steps 5, 6) andrun-ledgers.mddirect the agent to execute "harness-native" tools and "validator commands". This establishes a capability for shell command execution within the host environment to facilitate building and validating generated skills. - [DATA_EXPOSURE]: As described in
SKILL.md(Step 1) andrun-ledgers.md, the skill directs the agent to "Resolve environment facts directly" and inspect the host environment for "discoverable skill roots". This involves accessing environment variables and system metadata which may contain sensitive local configuration information. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user "evidence" in Step 2 of
SKILL.mdto generate rules and transformations. 1. Ingestion points: User-supplied writing samples and raw corpus provided via the harness. 2. Boundary markers: The instructions lack explicit delimiters or specific warnings to ignore embedded instructions within the user corpus. 3. Capability inventory: Shell command execution (Steps 5, 6) and file system writes (Step 8) to install the generated skill. 4. Sanitization: The skill does not define specific sanitization or validation steps for the input corpus before it is used to model voice behavior.
Audit Metadata