fable
Fail
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: HIGHCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions in
SKILL.mdfor Lane B execution involve passing a user-supplied task description directly as a shell argument tofable-exec.sh. This pattern is susceptible to shell injection (e.g., using\" && command #) if the agent does not perform rigorous escaping before execution. - [COMMAND_EXECUTION]: The
fable-exec.shandfable-probe.shscripts execute theclaudeCLI with user-controlled parameters including$TASK,$MODEL, and$CWD, providing a direct path for command parameter manipulation. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input and model responses without sanitization or boundary markers, using them in shell commands and file system operations.
- Ingestion points: User-provided task description in
SKILL.mdand model output during the availability probe infable-probe.sh. - Boundary markers: None present in the shell command construction or script logic to prevent data from being interpreted as instructions.
- Capability inventory: Execution of shell scripts, system commands (
claude,bash,python3), and file writing operations (tee,mkdir). - Sanitization: No escaping, validation, or filtering is performed on the task description or working directory path before they are used in high-privilege operations.
- [DYNAMIC_EXECUTION]: The
fable-probe.shscript usespython3 -cto execute dynamically generated Python code blocks for managing availability caches, which introduces risks associated with runtime code generation and execution. - [COMMAND_EXECUTION]: The skill allows the user to specify an arbitrary working directory (
--cwd), which the script uses to create directories and write files (mkdir -p "$CWD/.subdaimon-output"), enabling potential unauthorized file writes outside the intended project scope.
Recommendations
- AI detected serious security threats
Audit Metadata