travel-cost-estimator

Pass

Audited by Gen Agent Trust Hub on Aug 30, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • Browser Automation and DOM Manipulation: The skill utilizes Playwright to automate interactions with corporate booking portals. This includes the use of page.evaluate to execute JavaScript within the browser context to set values for hidden input fields. While this involves dynamic execution, it is a specific technique used to interact with modern web frameworks like React or Angular that often ignore standard value assignments.
  • Access to Sensitive Data Sources: The skill is designed to read information from two sensitive sources: a corporate booking tool (via an active browser session) and the user's mailbox (to retrieve expense report totals). This data is used to generate accurate travel estimates and benchmark non-bookable expenses. Review of these access patterns is recommended to ensure they align with organizational data handling policies.
  • Credential Handling Best Practices: The skill instructions explicitly forbid the agent from asking for user credentials. It relies entirely on existing Single Sign-On (SSO) sessions, which is a secure approach to accessing corporate resources without handling or storing passwords.
  • Indirect Prompt Injection Surface: By reading the user's mailbox for expense report summaries, the skill introduces a surface for indirect prompt injection. A malicious email formatted to look like an expense report could potentially influence the agent's behavior.
  • Ingestion points: The user's mailbox for expense system notification messages (Step 4).
  • Boundary markers: The skill uses explicit labels for data (e.g., labeling benchmarked data as "assumption") but does not define strict delimiters for the raw email content.
  • Capability inventory: The agent can automate browser actions, write to files, and send messages via Microsoft Teams.
  • Sanitization: The instructions focus on extracting specific totals and report IDs, which naturally limits the impact of embedded text, though explicit sanitization is not detailed.
  • Automated Messaging: The skill can draft and send messages to managers via Microsoft Teams. The instructions require explicit user confirmation and the display of the exact message text before any communication is sent, providing a human-in-the-loop safety checkpoint.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 30, 2026, 11:55 PM
Security Audit — agent-trust-hub — travel-cost-estimator