pp-table-reservation-goat

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill implements an 'Automatic Learning' and 'Playbook' system that ingests untrusted data to influence future agent behavior.
  • Ingestion points: Data enters the agent context through external reservation network APIs (OpenTable, Tock, Resy), user-provided queries passed to recall, and stored 'learnings' and 'playbooks'.
  • Boundary markers: The instructions mention a 'PII rule' for stripping personal identifiers and 'playbook notes' for documenting gotchas, but lacks strict delimiters to prevent the agent from obeying instructions embedded in restaurant data or saved playbooks.
  • Capability inventory: The skill has the capability to execute a wide range of commands through the table-reservation-goat-pp-cli binary, including booking and cancelling reservations.
  • Sanitization: The documentation advises 'PII discipline' to strip identifiers from queries and notes before teaching, though this is a manual process for the agent.
  • [EXTERNAL_DOWNLOADS]: The skill instructions direct the installation of CLI tools from the author's infrastructure.
  • Evidence: Fetches the installer from the @mvanhorn npm scope and the source code from the github.com/mvanhorn repository.
  • [DATA_EXFILTRATION]: The CLI tool includes a built-in feature to route output to external network sinks.
  • Evidence: The --deliver webhook:<url> flag allows the agent to POST the output of any command (which may include reservation details or authenticated profile data) to an arbitrary external URL.
  • [DYNAMIC_EXECUTION]: The skill uses a playbook system to dynamically generate and execute command sequences.
  • Evidence: The 'Playbook' feature replays sequences of steps with variable substitution (Playbook.steps) derived from previous sessions or 'preseeded' patterns. The agent is instructed to follow 'next_action' steps suggested by the CLI's internal learning engine verbatim.
  • [COMMAND_EXECUTION]: The skill involves the execution of local binaries and shell-based installation commands.
  • Evidence: Commands include npx, go install, and direct calls to table-reservation-goat-pp-cli.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 03:05 PM