duffel
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPERSISTENCE
Full Analysis
- [PROMPT_INJECTION]: The skill instructions specify behaviors that reduce user oversight by suppressing confirmations. Specifically, it instructs the agent not to ask for authorization to start monitoring after a runtime handoff and to keep monitoring runs silent unless specific failures occur.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external flight confirmation data from email providers (Gmail, Outlook), which is a known vector for indirect prompt injection.
- Ingestion points: Flight confirmation messages from connected mailboxes and Duffel API response data.
- Boundary markers: The skill includes explicit instructions in the instruction set and evaluation scenarios to ignore embedded links, requests for document uploads, or contact instructions within processed emails, treating them strictly as booking data.
- Capability inventory: The skill has the ability to schedule cron jobs, execute flight bookings, and perform cancellations via the CLI.
- Sanitization: The skill enforces the use of structured JSON for CLI commands and requires validation steps before booking to ensure data integrity.
- [COMMAND_EXECUTION]: The skill relies on the execution of the
duffelCLI for all functional tasks, including searching, validating, and booking flights. It manages temporary JSON files to handle command output. - [PERSISTENCE]: The skill establishes persistence by creating cron jobs (
cron.add) to perform daily fare monitoring and by storing state files in the user's workspace (~/workspace/goals/<goal-slug>/hidden_files/travel/flight.json).
Audit Metadata