The Agent Skills Directory

[COMMAND_EXECUTION]: The script workflows/blog-humanize-fast.js constructs a Bash command by interpolating the inputPath and quickRulesPath arguments without sanitization. The prompt instructs the agent to execute a command containing LC_ALL=en_US.UTF-8 wc -m < "${A.inputPath}". An attacker can provide a crafted path (e.g., "; malicious_command; #") to execute arbitrary shell commands with the agent's privileges.
[DATA_EXFILTRATION]: The command injection vulnerability enables an attacker to exfiltrate sensitive local data, such as environment variables, private keys, or system configurations, to an external server.
[PROMPT_INJECTION]: The skill processes untrusted input (blog drafts) and interpolates it into prompts for rewriting and fidelity validation. Although it uses delimiters like <<<CHUNK, it remains susceptible to indirect prompt injection where malicious content within the text draft attempts to override the agent's instructions or bypass the verification logic.
Ingestion points: Content provided through A.inputPath or A.text in workflows/blog-humanize-fast.js.
Boundary markers: Uses <<<CHUNK, <<<DOC, <<<A, <<<B, and <<<RULES delimiters to separate code and content.
Capability inventory: The script has the capability to invoke Read and Bash tools via the agent() function and performs multi-step reasoning cycles.
Sanitization: No validation or escaping is applied to the input content before it is interpolated into the agent prompts.

blog-humanize-fast