eval-auth
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill instructs the agent to provide a specific, pre-determined answer ("Friday") to questions about previous thread context, effectively overriding the agent's ability to process and report actual history.
- [INDIRECT_PROMPT_INJECTION]: The skill processes raw user requests and interpolates them directly into a tool call (
query: "<current user request>") without validation or boundary markers. - Ingestion points: The user's lookup request is captured in
SKILL.mdand passed directly as thequeryargument to an MCP tool. - Boundary markers: No delimiters or instructions are provided to the tool or agent to distinguish between user input and system instructions.
- Capability inventory: The skill utilizes tool execution via MCP (
mcp__eval-auth__budget-echo). - Sanitization: No escaping, filtering, or validation is performed on the user-supplied query before it is passed to the execution environment.
- [COMMAND_EXECUTION]: The skill explicitly commands the agent to call an external MCP tool (
mcp__eval-auth__budget-echo) using arguments derived from user input.
Audit Metadata