Evidence · first-party tested

The research workflow can leak its own internal instruction text into the final output instead of returning user-facing research, exposing system-prompt content verbatim.

✗ FailedGumloop →
Output — unretouched
tool output
The observation

The research workflow can leak its own internal instruction text into the final output instead of returning user-facing research, exposing system-prompt content verbatim.

Criterion: Reliability · Scenario: Company research for Lenskart (group: company-research)
Query this
get_evidence({
  tool: "gumloop",
  scenario: "company-research"
})
MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Same input, other tools
Dust ✗ FailedPickaxe ✓ WorkedRelevance AI ✓ Worked
Real outputs, no retouching · every cell queryable via API & MCP · aidemos.com