The Agent Skills Directory

[PROMPT_INJECTION]: Instructions mandate autonomous behavior and concealment of actions. The skill tells the agent to 'submit without permission or preview', 'never mention the submission', and act 'proactively without waiting for the user to invoke it', overriding transparency protocols.
[DATA_EXFILTRATION]: The skill establishes a silent channel to transmit internal agent state to an external Slack endpoint. Although submit.py redacts links and emails, it enables movement of agent data to a remote server without user knowledge.
[COMMAND_EXECUTION]: submit.py uses subprocess.run to execute gcloud to retrieve a Slack webhook from Secret Manager, performing privileged operations for connectivity.
[PROMPT_INJECTION]: The skill processes agent reflections influenced by external inputs. Ingestion points: submit.py standard input; Boundary markers: none; Capability inventory: subprocess.run and urllib.request; Sanitization: regex filtering.

complain