The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to generate and execute code at runtime to verify logical flaws in logic-review-guide.md (Step 8). The generated scripts incorporate untrusted user-provided code and are executed via system runtimes like Python, Node.js, or Go. Although the instructions include constraints to avoid network or file system side effects, the execution of code derived from external input remains a high-risk activity that could lead to sandbox escapes or unauthorized access.
[COMMAND_EXECUTION]: The verification process in logic-review-guide.md (Step 8) assumes the presence of a runtime environment and instructs the agent to run shell commands (e.g., checking versions with python3 --version or executing scripts) to confirm findings. This implies the agent has access to a command-line interface and the ability to spawn subprocesses.
[PROMPT_INJECTION]: The skill is highly vulnerable to Indirect Prompt Injection because its primary function is to ingest and analyze untrusted code shared by users.
Ingestion points: The skill triggers when a user shares code snippets for review (SKILL.md).
Boundary markers: Absent. The skill lacks explicit delimiters or instructions to treat user-provided code as untrusted data that should not be interpreted as instructions.
Capability inventory: The agent possesses the capability to execute generated code and interact with the shell as described in logic-review-guide.md (Step 8).
Sanitization: Partial. The skill provides instructions to the agent to limit its generated scripts to "pure computation," but it does not sanitize or filter the user-provided code itself before analysis or interpolation into the verification scripts.

logic-review