The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/ddmin.py provides a mechanism to execute arbitrary programs as part of its 'oracle' functionality. It utilizes subprocess.run() to invoke commands provided via command-line arguments. While the script avoids shell execution by passing arguments as a list (shell=False), it is designed to run external binaries (e.g., compilers, parsers, or test scripts) repeatedly against varying inputs.
Evidence: scripts/ddmin.py (lines 67-75) constructs a command vector argv from user-supplied oracle_words and executes it using subprocess.run with a 60-second timeout.
[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted data files to reduce failing inputs. This content is then passed to an external 'oracle' command for evaluation. If the input data contains malicious instructions that the oracle program (or the agent interpreting the oracle's output) obeys, it could lead to indirect injection attacks.
Ingestion points: scripts/ddmin.py (lines 142-144) reads the entire content of a file provided in the input argument.
Boundary markers: The skill and its associated scripts do not implement delimiters or 'ignore instructions' warnings for the content being processed.
Capability inventory: scripts/ddmin.py has the capability to execute shell commands via subprocess.run and write to the file system using tempfile.NamedTemporaryFile.
Sanitization: No sanitization, escaping, or validation is performed on the input file's content before it is written to a temporary file and passed to the oracle command for execution.

write-behavior-tests