The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the execution of a local shell script (probe.sh) to perform its core tasks. This script manages temporary directories and invokes the claude command-line interface multiple times in parallel.
Evidence: bash .agents/skills/term-check/probe.sh <scenario-file> <term-a> <term-b> <term-c>
Potential Vulnerability: In probe.sh, the command sed "s|{TERM}|$term|g" "$SCENARIO" uses a pipe (|) as a delimiter. A malicious or malformed term containing the pipe character could disrupt the sed operation, leading to script failure or unexpected prompt construction.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from a user-provided scenario file and interpolates terms into it, which are then passed to an LLM without sanitization.
Ingestion points: The script probe.sh reads input from the file path provided in the first argument (<scenario-file>) and the subsequent term arguments.
Boundary markers: None. The resulting text from the sed operation is passed directly as a prompt to the claude CLI.
Capability inventory: The skill can execute bash scripts, perform file system operations via sed and mktemp, and invoke the agent's CLI tool (claude) with low effort and strict MCP configuration.
Sanitization: There is no evidence of validation or escaping for the content of the scenario file or the terms before they are interpolated and evaluated by the models.

term-check