The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies that task descriptions (beans) are untrusted data sources that could potentially contain malicious instructions. It mitigates this risk by requiring an explicit approval workflow for 'runnable gates,' ensuring that the agent inspects and validates every command against a safety checklist before execution.
[COMMAND_EXECUTION]: While the skill involves executing commands for development and verification (such as make test, curl, and kubectl), it emphasizes that these should only be run after manual inspection and in the correct context. The examples provided are standard for software development and do not exhibit malicious patterns.
[DATA_EXFILTRATION]: The instructions explicitly guide the agent to avoid running commands that access secrets or sensitive environment variables. It also recommends using fingerprints (hashes) of command output instead of persisting raw output, which reduces the risk of accidental sensitive data exposure in logs.

keep-on-going