The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the user to install binary tools using npx -y @mvanhorn/printing-press-library and go install github.com/mvanhorn/.... These resources are owned by the vendor 'mvanhorn' and are considered safe under vendor trust rules, but represent remote code execution at installation time.
[DATA_EXFILTRATION]: The CLI supports a --deliver webhook:<url> flag, which POSTs command output to an external endpoint. This feature could be leveraged to exfiltrate sensitive data if the sink URL is controlled by an attacker.
[DYNAMIC_EXECUTION]: The skill implements a 'learning loop' where the agent is instructed to retrieve and execute 'playbooks' (sequences of shell commands) from a local store via the recall command. The execution of these stored strings constitutes dynamic execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests machine-hostile weekly price files from the Pakistan Bureau of Statistics and uses a teach mechanism to persist mappings between queries and resources. A lack of sanitization could allow malicious content in the external source or a poisoned teach command to influence future automated 'playbook' actions.
Ingestion points: Parses external report files and index pages; accepts user-provided queries for the teach and recall commands.
Boundary markers: None identified in the instruction text or command logic.
Capability inventory: Executes shell commands, performs network POST requests via webhooks, and writes files via the file:<path> sink.
Sanitization: No specific sanitization or validation of external cell values or stored playbooks is described.
[COMMAND_EXECUTION]: The skill involves the direct execution of the pbs-pp-cli binary with various arguments and uses npx and go for environment setup.

pp-pbs