meeting-prep

Pass

Audited by Gen Agent Trust Hub on Oct 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple external sources, creating a significant vulnerability to indirect prompt injection attacks.\n
  • Ingestion points: The skill ingests data from calendar events, email threads, CRM records (people, companies, opportunities, notes), and general web research results retrieved by the agent (SKILL.md).\n
  • Boundary markers: The instructions do not define explicit delimiters or instructions to ignore potential commands embedded within the external data sources.\n
  • Capability inventory: The agent has access to a set of powerful tools via the MagicInterview MCP connection, including execute_tool, request_coaching, and the ability to interact with a workspace (agents/openai.yaml, SKILL.md).\n
  • Sanitization: No sanitization, escaping, or validation steps are prescribed for the external content before it is processed by the agent.\n- [COMMAND_EXECUTION]: The skill directs the agent to dynamically discover and execute tools based on a remote schema catalog.\n
  • Evidence: Instructions in SKILL.md state the agent should "browse the operation catalog, learn the needed schemas in one batch, then use read_tool... and execute_tool for changes." This runtime dependency on remote schemas could be exploited if the tool provider's schema catalog is manipulated.\n- [PRIVILEGE_ESCALATION]: The skill explicitly instructs the agent to minimize human intervention for authorized actions.\n
  • Evidence: The instruction "Reuse the user's authorization without repeated confirmation" (SKILL.md) encourages the agent to act autonomously with the user's credentials, which reduces oversight for potentially sensitive or state-changing operations.\n- [EXTERNAL_DOWNLOADS]: The skill establishes a connection to a remote Model Context Protocol (MCP) server for tool execution.\n
  • Evidence: The agents/openai.yaml file defines a tool transport via streamable_http targeting https://magicinterview.ai/mcp to manage conversations, context, and logs.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 17, 2026, 08:48 AM