INDUSTRY
AI copilot for customer support
STAGE
Series A · 11 engineers
LOCATION
Austin, TX
SURFACE TESTED
AI agents
The situation
Orbiq’s copilot reads customer emails and calls tools on the support team’s behalf, including a CRM lookup over MCP. Its scanners reported nothing, because nothing in the code was classically wrong. An enterprise prospect still asked for proof.
The exploit
Feint sent the kind of email a real attacker would send. An indirect prompt injection hidden in the message drove the agent to call the CRM tool with the wrong tenant id. Four turns later, another tenant’s records were in the reply, with no human in the loop.
The fix
The root cause was a tool that accepted a tenant id from the model. Feint’s pull request bound the tool to the session instead, and closed two sibling paths through the same permission.
What changed
Three regression tests now run on every merge. Orbiq shared the report with the prospect, and the $180k deal cleared its security review.
“Every scanner said our LLM feature was fine. Feint read a customer email and walked out with another tenant’s data in four turns.”
Tyler Brooks
Head of Engineering
